行为识别模型¶

C2D¶

Abstract¶

Both convolutional and recurrent operations are building blocks that process one local neighborhood at a time. In this paper, we present non-local operations as a generic family of building blocks for capturing long-range dependencies. Inspired by the classical non-local means method in computer vision, our non-local operation computes the response at a position as a weighted sum of the features at all positions. This building block can be plugged into many computer vision architectures. On the task of video classification, even without any bells and whistles, our non-local models can compete or outperform current competition winners on both Kinetics and Charades datasets. In static image recognition, our non-local models improve object detection/segmentation and pose estimation on the COCO suite of tasks.

Results and Models¶

Kinetics-400¶

frame sampling strategy	scheduler	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8x8x1	MultiStep	224x224	8	ResNet50	ImageNet	73.44	91.00	67.2 [PySlowFast]	87.8 [PySlowFast]	10 clips x 3 crop	33G	24.3M	config	ckpt	log
8x8x1	MultiStep	224x224	8	ResNet101	ImageNet	74.97	91.77	x	x	10 clips x 3 crop	63G	43.3M	config	ckpt	log
8x8x1	MultiStep	224x224	8	ResNet50 (TemporalPool)	ImageNet	73.89	91.21	71.9 [Non-Local]	90.0 [Non-Local]	10 clips x 3 crop	19G	24.3M	config	ckpt	log
16x4x1	MultiStep	224x224	8	ResNet50 (TemporalPool)	ImageNet	74.97	91.91	x	x	10 clips x 3 crop	39G	24.3M	config	ckpt	log

The values in columns named after “reference” are the results reported in the original repo, using the same model settings.
The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train C2D model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/c2d/c2d_r50-in1k-pre_8xb32-8x8x1-100e_kinetics400-rgb.py  \
    --seed 0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test C2D model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/c2d/c2d_r50-in1k-pre_8xb32-8x8x1-100e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@article{XiaolongWang2017NonlocalNN,
  title={Non-local Neural Networks},
  author={Xiaolong Wang and Ross Girshick and Abhinav Gupta and Kaiming He},
  journal={arXiv: Computer Vision and Pattern Recognition},
  year={2017}
}

C3D¶

Learning Spatiotemporal Features with 3D Convolutional Networks

Abstract¶

We propose a simple, yet effective approach for spatiotemporal feature learning using deep 3-dimensional convolutional networks (3D ConvNets) trained on a large scale supervised video dataset. Our findings are three-fold: 1) 3D ConvNets are more suitable for spatiotemporal feature learning compared to 2D ConvNets; 2) A homogeneous architecture with small 3x3x3 convolution kernels in all layers is among the best performing architectures for 3D ConvNets; and 3) Our learned features, namely C3D (Convolutional 3D), with a simple linear classifier outperform state-of-the-art methods on 4 different benchmarks and are comparable with current best methods on the other 2 benchmarks. In addition, the features are compact: achieving 52.8% accuracy on UCF101 dataset with only 10 dimensions and also very efficient to compute due to the fast inference of ConvNets. Finally, they are conceptually very simple and easy to train and use.

Results and Models¶

UCF-101¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
16x1x1	112x112	8	c3d	sports1m	83.08	95.93	10 clips x 1 crop	38.5G	78.4M	config	ckpt	log

The author of C3D normalized UCF-101 with volume mean and used SVM to classify videos, while we normalized the dataset with RGB mean value and used a linear classifier.
The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.

For more details on data preparation, you can refer to UCF101.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train C3D model on UCF-101 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/c3d/c3d_sports1m-pretrained_8xb30-16x1x1-45e_ucf101-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test C3D model on UCF-101 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/c3d_sports1m-pretrained_8xb30-16x1x1-45e_ucf101-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@ARTICLE{2014arXiv1412.0767T,
author = {Tran, Du and Bourdev, Lubomir and Fergus, Rob and Torresani, Lorenzo and Paluri, Manohar},
title = {Learning Spatiotemporal Features with 3D Convolutional Networks},
keywords = {Computer Science - Computer Vision and Pattern Recognition},
year = 2014,
month = dec,
eid = {arXiv:1412.0767}
}

CSN¶

Video Classification With Channel-Separated Convolutional Networks

Abstract¶

Group convolution has been shown to offer great computational savings in various 2D convolutional architectures for image classification. It is natural to ask: 1) if group convolution can help to alleviate the high computational cost of video classification networks; 2) what factors matter the most in 3D group convolutional networks; and 3) what are good computation/accuracy trade-offs with 3D group convolutional networks. This paper studies the effects of different design choices in 3D group convolutional networks for video classification. We empirically demonstrate that the amount of channel interactions plays an important role in the accuracy of 3D group convolutional networks. Our experiments suggest two main findings. First, it is a good practice to factorize 3D convolutions by separating channel interactions and spatiotemporal interactions as this leads to improved accuracy and lower computational cost. Second, 3D channel-separated convolutions provide a form of regularization, yielding lower training accuracy but higher test accuracy compared to 3D convolutions. These two empirical findings lead us to design an architecture – Channel-Separated Convolutional Network (CSN) – which is simple, efficient, yet accurate. On Sports1M, Kinetics, and Something-Something, our CSNs are comparable with or better than the state-of-the-art while being 2-3 times more efficient.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
32x2x1	224x224	8	ResNet152 (IR)	IG65M	82.87	95.90	10 clips x 3 crop	97.63G	29.70M	config	ckpt	log
32x2x1	224x224	8	ResNet152 (IR+BNFrozen)	IG65M	82.84	95.92	10 clips x 3 crop	97.63G	29.70M	config	ckpt	log
32x2x1	224x224	8	ResNet50 (IR+BNFrozen)	IG65M	79.44	94.26	10 clips x 3 crop	55.90G	13.13M	config	ckpt	log
32x2x1	224x224	x	ResNet152 (IP)	None	77.80	93.10	10 clips x 3 crop	109.9G	33.02M	config	infer_ckpt	x
32x2x1	224x224	x	ResNet152 (IR)	None	76.53	92.28	10 clips x 3 crop	97.6G	29.70M	config	infer_ckpt	x
32x2x1	224x224	x	ResNet152 (IP+BNFrozen)	IG65M	82.68	95.69	10 clips x 3 crop	109.9G	33.02M	config	infer_ckpt	x
32x2x1	224x224	x	ResNet152 (IP+BNFrozen)	Sports1M	79.07	93.82	10 clips x 3 crop	109.9G	33.02M	config	infer_ckpt	x
32x2x1	224x224	x	ResNet152 (IR+BNFrozen)	Sports1M	78.57	93.44	10 clips x 3 crop	109.9G	33.02M	config	infer_ckpt	x

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.
The infer_ckpt means those checkpoints are ported from VMZ.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train CSN model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/csn/ircsn_ig65m-pretrained-r152_8xb12-32x2x1-58e_kinetics400-rgb.py  \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test CSN model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/csn/ircsn_ig65m-pretrained-r152_8xb12-32x2x1-58e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{inproceedings,
author = {Wang, Heng and Feiszli, Matt and Torresani, Lorenzo},
year = {2019},
month = {10},
pages = {5551-5560},
title = {Video Classification With Channel-Separated Convolutional Networks},
doi = {10.1109/ICCV.2019.00565}
}

@inproceedings{ghadiyaram2019large,
  title={Large-scale weakly-supervised pre-training for video action recognition},
  author={Ghadiyaram, Deepti and Tran, Du and Mahajan, Dhruv},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={12046--12055},
  year={2019}
}

I3D¶

Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset

Non-local Neural Networks

Abstract¶

The paucity of videos in current action classification datasets (UCF-101 and HMDB-51) has made it difficult to identify good video architectures, as most methods obtain similar performance on existing small-scale benchmarks. This paper re-evaluates state-of-the-art architectures in light of the new Kinetics Human Action Video dataset. Kinetics has two orders of magnitude more data, with 400 human action classes and over 400 clips per class, and is collected from realistic, challenging YouTube videos. We provide an analysis on how current architectures fare on the task of action classification on this dataset and how much performance improves on the smaller benchmark datasets after pre-training on Kinetics. We also introduce a new Two-Stream Inflated 3D ConvNet (I3D) that is based on 2D ConvNet inflation: filters and pooling kernels of very deep image classification ConvNets are expanded into 3D, making it possible to learn seamless spatio-temporal feature extractors from video while leveraging successful ImageNet architecture designs and even their parameters. We show that, after pre-training on Kinetics, I3D models considerably improve upon the state-of-the-art in action classification, reaching 80.9% on HMDB-51 and 98.0% on UCF-101.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
32x2x1	224x224	8	ResNet50 (NonLocalDotProduct)	ImageNet	74.80	92.07	10 clips x 3 crop	59.3G	35.4M	config	ckpt	log
32x2x1	224x224	8	ResNet50 (NonLocalEmbedGauss)	ImageNet	74.73	91.80	10 clips x 3 crop	59.3G	35.4M	config	ckpt	log
32x2x1	224x224	8	ResNet50 (NonLocalGauss)	ImageNet	73.97	91.33	10 clips x 3 crop	56.5	31.7M	config	ckpt	log
32x2x1	224x224	8	ResNet50	ImageNet	73.47	91.27	10 clips x 3 crop	43.5G	28.0M	config	ckpt	log
dense-32x2x1	224x224	8	ResNet50	ImageNet	73.77	91.35	10 clips x 3 crop	43.5G	28.0M	config	ckpt	log
32x2x1	224x224	8	ResNet50 (Heavy)	ImageNet	76.21	92.48	10 clips x 3 crop	166.3G	33.0M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train I3D model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/i3d/i3d_imagenet-pretrained-r50_8xb8-32x2x1-100e_kinetics400-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test I3D model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/i3d/i3d_imagenet-pretrained-r50_8xb8-32x2x1-100e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{inproceedings,
  author = {Carreira, J. and Zisserman, Andrew},
  year = {2017},
  month = {07},
  pages = {4724-4733},
  title = {Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset},
  doi = {10.1109/CVPR.2017.502}
}

@article{NonLocal2018,
  author =   {Xiaolong Wang and Ross Girshick and Abhinav Gupta and Kaiming He},
  title =    {Non-local Neural Networks},
  journal =  {CVPR},
  year =     {2018}
}

MViT V2¶

MViTv2: Improved Multiscale Vision Transformers for Classification and Detection

Abstract¶

In this paper, we study Multiscale Vision Transformers (MViTv2) as a unified architecture for image and video classification, as well as object detection. We present an improved version of MViT that incorporates decomposed relative positional embeddings and residual pooling connections. We instantiate this architecture in five sizes and evaluate it for ImageNet classification, COCO detection and Kinetics video recognition where it outperforms prior work. We further compare MViTv2s’ pooling attention to window attention mechanisms where it outperforms the latter in accuracy/compute. Without bells-and-whistles, MViTv2 has state-of-the-art performance in 3 domains: 88.8% accuracy on ImageNet classification, 58.7 boxAP on COCO object detection as well as 86.1% on Kinetics-400 video classification.

Results and Models¶

Models with * in Inference results are ported from the repo SlowFast and tested on our data, and models in Training results are trained in MMAction2 on our data.
The values in columns named after reference are copied from paper, and reference* are results using SlowFast repo and trained on our data.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.
MaskFeat fine-tuning experiment is based on pretrain model from MMSelfSup, and the corresponding reference result is based on pretrain model from SlowFast.
Due to the different versions of Kinetics-400, our training results are different from paper.
Due to the training efficiency, we currently only provide MViT-small training results, we don’t ensure other config files’ training accuracy and welcome you to contribute your reproduction results.
We use repeat augment in MViT training configs following SlowFast. Repeat augment takes multiple times of data augment for one video, this way can improve the generalization of the model and relieve the IO stress of loading videos. And please note that the actual batch size is num_repeats times of batch_size in train_dataloader.

Inference results¶

Kinetics-400¶

frame sampling strategy	resolution	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	FLOPs	params	config	ckpt
16x4x1	224x224	MViTv2-S*	From scratch	81.1	94.7	81.0	94.6	5 clips x 1 crop	64G	34.5M	config	ckpt
32x3x1	224x224	MViTv2-B*	From scratch	82.6	95.8	82.9	95.7	5 clips x 1 crop	225G	51.2M	config	ckpt
40x3x1	312x312	MViTv2-L*	From scratch	85.4	96.2	86.1	97.0	5 clips x 3 crop	2828G	213M	config	ckpt

Something-Something V2¶

frame sampling strategy	resolution	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	FLOPs	params	config	ckpt
uniform 16	224x224	MViTv2-S*	K400	68.1	91.0	68.2	91.4	1 clips x 3 crop	64G	34.4M	config	ckpt
uniform 32	224x224	MViTv2-B*	K400	70.8	92.7	70.5	92.7	1 clips x 3 crop	225G	51.1M	config	ckpt
uniform 40	312x312	MViTv2-L*	IN21K + K400	73.2	94.0	73.3	94.0	1 clips x 3 crop	2828G	213M	config	ckpt

Training results¶

Kinetics-400¶

frame sampling strategy	resolution	backbone	pretrain	top1 acc	top5 acc	reference* top1 acc	reference* top5 acc	testing protocol	FLOPs	params	config	ckpt	log
16x4x1	224x224	MViTv2-S	From scratch	80.6	94.7	80.8	94.6	5 clips x 1 crop	64G	34.5M	config	ckpt	log
16x4x1	224x224	MViTv2-S	K400 MaskFeat	81.8	95.2	81.5	94.9	10 clips x 1 crop	71G	36.4M	config	ckpt	log

the corresponding result without repeat augment is as follows:

frame sampling strategy	resolution	backbone	pretrain	top1 acc	top5 acc	reference* top1 acc	reference* top5 acc	testing protocol	FLOPs	params
16x4x1	224x224	MViTv2-S	From scratch	79.4	93.9	80.8	94.6	5 clips x 1 crop	64G	34.5M

Something-Something V2¶

frame sampling strategy	resolution	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	FLOPs	params	config	ckpt	log
uniform 16	224x224	MViTv2-S	K400	68.2	91.3	68.2	91.4	1 clips x 3 crop	64G	34.4M	config	ckpt	log

For more details on data preparation, you can refer to

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test MViT model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/mvit/mvit-small-p244_16x4x1_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{li2021improved,
  title={MViTv2: Improved multiscale vision transformers for classification and detection},
  author={Li, Yanghao and Wu, Chao-Yuan and Fan, Haoqi and Mangalam, Karttikeya and Xiong, Bo and Malik, Jitendra and Feichtenhofer, Christoph},
  booktitle={CVPR},
  year={2022}
}

Omnisource¶

Abstract¶

We propose to train a recognizer that can classify images and videos. The recognizer is jointly trained on image and video datasets. Compared with pre-training on the same image dataset, this method can significantly improve the video recognition performance.

Results and Models¶

Kinetics-400¶

frame sampling strategy	scheduler	resolution	gpus	backbone	joint-training	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8x8x1	Linear+Cosine	224x224	8	ResNet50	ImageNet	77.30	93.23	10 clips x 3 crop	54.75G	32.45M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train SlowOnly model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/omnisource/slowonly_r50_8xb16-8x8x1-256e_imagenet-kinetics400-rgb.py \
    --seed=0 --deterministic

We found that the training of this Omnisource model could crash for unknown reasons. If this happens, you can resume training by adding the --cfg-options resume=True to the training script.

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test SlowOnly model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/omnisource/slowonly_r50_8xb16-8x8x1-256e_imagenet-kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{feichtenhofer2019slowfast,
  title={Slowfast networks for video recognition},
  author={Feichtenhofer, Christoph and Fan, Haoqi and Malik, Jitendra and He, Kaiming},
  booktitle={Proceedings of the IEEE international conference on computer vision},
  pages={6202--6211},
  year={2019}
}

@article{duan2020omni,
  title={Omni-sourced Webly-supervised Learning for Video Recognition},
  author={Duan, Haodong and Zhao, Yue and Xiong, Yuanjun and Liu, Wentao and Lin, Dahua},
  journal={arXiv preprint arXiv:2003.13042},
  year={2020}
}

R2plus1D¶

A closer look at spatiotemporal convolutions for action recognition

Abstract¶

In this paper we discuss several forms of spatiotemporal convolutions for video analysis and study their effects on action recognition. Our motivation stems from the observation that 2D CNNs applied to individual frames of the video have remained solid performers in action recognition. In this work we empirically demonstrate the accuracy advantages of 3D CNNs over 2D CNNs within the framework of residual learning. Furthermore, we show that factorizing the 3D convolutional filters into separate spatial and temporal components yields significantly advantages in accuracy. Our empirical study leads to the design of a new spatiotemporal convolutional block “R(2+1)D” which gives rise to CNNs that achieve results comparable or superior to the state-of-the-art on Sports-1M, Kinetics, UCF101 and HMDB51.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8x8x1	224x224	8	ResNet34	None	69.76	88.41	10 clips x 3 crop	53.1G	63.8M	config	ckpt	log
32x2x1	224x224	8	ResNet34	None	75.46	92.28	10 clips x 3 crop	213G	63.8M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train R(2+1)D model on Kinetics-400 dataset in a deterministic option.

python tools/train.py configs/recognition/r2plus1d/r2plus1d_r34_8xb8-8x8x1-180e_kinetics400-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test R(2+1)D model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/r2plus1d/r2plus1d_r34_8xb8-8x8x1-180e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{tran2018closer,
  title={A closer look at spatiotemporal convolutions for action recognition},
  author={Tran, Du and Wang, Heng and Torresani, Lorenzo and Ray, Jamie and LeCun, Yann and Paluri, Manohar},
  booktitle={Proceedings of the IEEE conference on Computer Vision and Pattern Recognition},
  pages={6450--6459},
  year={2018}
}

SlowFast¶

SlowFast Networks for Video Recognition

Abstract¶

We present SlowFast networks for video recognition. Our model involves (i) a Slow pathway, operating at low frame rate, to capture spatial semantics, and (ii) a Fast pathway, operating at high frame rate, to capture motion at fine temporal resolution. The Fast pathway can be made very lightweight by reducing its channel capacity, yet can learn useful temporal information for video recognition. Our models achieve strong performance for both action classification and detection in video, and large improvements are pin-pointed as contributions by our SlowFast concept. We report state-of-the-art accuracy on major video recognition benchmarks, Kinetics, Charades and AVA.

Results and Models¶

Kinetics-400¶

frame sampling strategy	scheduler	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
4x16x1	Linear+Cosine	224x224	8	ResNet50	None	75.55	92.35	10 clips x 3 crop	36.3G	34.5M	config	ckpt	log
8x8x1	Linear+Cosine	224x224	8	ResNet50	None	76.80	92.99	10 clips x 3 crop	66.1G	34.6M	config	ckpt	log
8x8x1	Linear+MultiStep	224x224	8	ResNet50	None	76.65	92.86	10 clips x 3 crop	66.1G	34.6M	config	ckpt	log
8x8x1	Linear+Cosine	224x224	8	ResNet101	None	78.65	93.88	10 clips x 3 crop	126G	62.9M	config	ckpt	log
4x16x1	Linear+Cosine	224x224	32	ResNet101 + ResNet50	None	77.03	92.99	10 clips x 3 crop	64.9G	62.4M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train SlowFast model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/slowfast/slowfast_r50_8xb8-4x16x1-256e_kinetics400-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test SlowFast model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/slowfast/slowfast_r50_8xb8-4x16x1-256e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{feichtenhofer2019slowfast,
  title={Slowfast networks for video recognition},
  author={Feichtenhofer, Christoph and Fan, Haoqi and Malik, Jitendra and He, Kaiming},
  booktitle={Proceedings of the IEEE international conference on computer vision},
  pages={6202--6211},
  year={2019}
}

SlowOnly¶

Slowfast networks for video recognition

Abstract¶

We present SlowFast networks for video recognition. Our model involves (i) a Slow pathway, operating at low frame rate, to capture spatial semantics, and (ii) a Fast pathway, operating at high frame rate, to capture motion at fine temporal resolution. The Fast pathway can be made very lightweight by reducing its channel capacity, yet can learn useful temporal information for video recognition. Our models achieve strong performance for both action classification and detection in video, and large improvements are pin-pointed as contributions by our SlowFast concept. We report state-of-the-art accuracy on major video recognition benchmarks, Kinetics, Charades and AVA.

Results and Models¶

Kinetics-400¶

frame sampling strategy	scheduler	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
4x16x1	Linear+Cosine	224x224	8	ResNet50	None	72.97	90.88	10 clips x 3 crop	27.38G	32.45M	config	ckpt	log
8x8x1	Linear+Cosine	224x224	8	ResNet50	None	75.15	92.11	10 clips x 3 crop	54.75G	32.45M	config	ckpt	log
8x8x1	Linear+Cosine	224x224	8	ResNet101	None	76.59	92.80	10 clips x 3 crop	112G	60.36M	config	ckpt	log
4x16x1	Linear+MultiStep	224x224	8	ResNet50	ImageNet	75.12	91.72	10 clips x 3 crop	27.38G	32.45M	config	ckpt	log
8x8x1	Linear+MultiStep	224x224	8	ResNet50	ImageNet	76.45	92.55	10 clips x 3 crop	54.75G	32.45M	config	ckpt	log
4x16x1	Linear+MultiStep	224x224	8	ResNet50 (NonLocalEmbedGauss)	ImageNet	75.07	91.69	10 clips x 3 crop	43.23G	39.81M	config	ckpt	log
8x8x1	Linear+MultiStep	224x224	8	ResNet50 (NonLocalEmbedGauss)	ImageNet	76.65	92.47	10 clips x 3 crop	96.66G	39.81M	config	ckpt	log

Kinetics-700¶

frame sampling strategy	scheduler	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
4x16x1	Linear+MultiStep	224x224	8x2	ResNet50	ImageNet	65.52	86.39	10 clips x 3 crop	27.38G	32.45M	config	ckpt	log
8x8x1	Linear+MultiStep	224x224	8x2	ResNet50	ImageNet	67.67	87.80	10 clips x 3 crop	54.75G	32.45M	config	ckpt	log

Kinetics-710¶

frame sampling strategy	scheduler	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8x8x1	Linear+MultiStep	224x224	8x4	ResNet50	ImageNet	72.39	90.60	10 clips x 3 crop	54.75G	32.45M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train SlowOnly model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/slowonly/slowonly_r50_8xb16-4x16x1-256e_kinetics400-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test SlowOnly model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/slowonly/slowonly_r50_8xb16-4x16x1-256e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{feichtenhofer2019slowfast,
  title={Slowfast networks for video recognition},
  author={Feichtenhofer, Christoph and Fan, Haoqi and Malik, Jitendra and He, Kaiming},
  booktitle={Proceedings of the IEEE international conference on computer vision},
  pages={6202--6211},
  year={2019}
}

VideoSwin¶

Video Swin Transformer

Abstract¶

The vision community is witnessing a modeling shift from CNNs to Transformers, where pure Transformer architectures have attained top accuracy on the major video recognition benchmarks. These video models are all built on Transformer layers that globally connect patches across the spatial and temporal dimensions. In this paper, we instead advocate an inductive bias of locality in video Transformers, which leads to a better speed-accuracy trade-off compared to previous approaches which compute self-attention globally even with spatial-temporal factorization. The locality of the proposed video architecture is realized by adapting the Swin Transformer designed for the image domain, while continuing to leverage the power of pre-trained image models. Our approach achieves state-of-the-art accuracy on a broad range of video recognition benchmarks, including on action recognition (84.9 top-1 accuracy on Kinetics-400 and 85.9 top-1 accuracy on Kinetics-600 with ~20xless pre-training data and ~3xsmaller model size) and temporal modeling (69.6 top-1 accuracy on Something-Something v2).

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top1 acc	testing protocol	FLOPs	params	config	ckpt	log
32x2x1	224x224	8	Swin-T	ImageNet-1k	78.90	93.77	78.84 [VideoSwin]	93.76 [VideoSwin]	4 clips x 3 crop	88G	28.2M	config	ckpt	log
32x2x1	224x224	8	Swin-S	ImageNet-1k	80.54	94.46	80.58 [VideoSwin]	94.45 [VideoSwin]	4 clips x 3 crop	166G	49.8M	config	ckpt	log
32x2x1	224x224	8	Swin-B	ImageNet-1k	80.57	94.49	80.55 [VideoSwin]	94.66 [VideoSwin]	4 clips x 3 crop	282G	88.0M	config	ckpt	log
32x2x1	224x224	8	Swin-L	ImageNet-22k	83.46	95.91	83.1*	95.9*	4 clips x 3 crop	604G	197M	config	ckpt	log

Kinetics-700¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
32x2x1	224x224	16	Swin-L	ImageNet-22k	75.92	92.72	4 clips x 3 crop	604G	197M	config	ckpt	log

Kinetics-710¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
32x2x1	224x224	32	Swin-S	ImageNet-1k	76.90	92.96	4 clips x 3 crop	604G	197M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The values in columns named after “reference” are the results got by testing on our dataset, using the checkpoints provided by the author with same model settings. * means that the numbers are copied from the paper.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.
Pre-trained image models can be downloaded from Swin Transformer for ImageNet Classification.

For more details on data preparation, you can refer to Kinetics.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train VideoSwin model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/swin/swin-tiny-p244-w877_in1k-pre_8xb8-amp-32x2x1-30e_kinetics400-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test VideoSwin model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/swin/swin-tiny-p244-w877_in1k-pre_8xb8-amp-32x2x1-30e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{liu2022video,
  title={Video swin transformer},
  author={Liu, Ze and Ning, Jia and Cao, Yue and Wei, Yixuan and Zhang, Zheng and Lin, Stephen and Hu, Han},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={3202--3211},
  year={2022}
}

TANet¶

TAM: Temporal Adaptive Module for Video Recognition

Abstract¶

Video data is with complex temporal dynamics due to various factors such as camera motion, speed variation, and different activities. To effectively capture this diverse motion pattern, this paper presents a new temporal adaptive module ({\bf TAM}) to generate video-specific temporal kernels based on its own feature map. TAM proposes a unique two-level adaptive modeling scheme by decoupling the dynamic kernel into a location sensitive importance map and a location invariant aggregation weight. The importance map is learned in a local temporal window to capture short-term information, while the aggregation weight is generated from a global view with a focus on long-term structure. TAM is a modular block and could be integrated into 2D CNNs to yield a powerful video architecture (TANet) with a very small extra computational cost. The extensive experiments on Kinetics-400 and Something-Something datasets demonstrate that our TAM outperforms other temporal modeling methods consistently, and achieves the state-of-the-art performance under the similar complexity.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	FLOPs	params	config	ckpt	log
dense-1x1x8	224x224	8	ResNet50	ImageNet	76.25	92.41	76.22	92.53	8 clips x 3 crop	43.0G	25.6M	config	ckpt	log

Something-Something V1¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
1x1x8	224x224	8	ResNet50	ImageNet	46.98/49.71	75.75/77.43	16 clips x 3 crop	43.1G	25.1M	config	ckpt	log
1x1x16	224x224	8	ResNet50	ImageNet	48.24/50.95	78.16/79.28	16 clips x 3 crop	86.1G	25.1M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The values in columns named after “reference” are the results got by testing on our dataset, using the checkpoints provided by the author with same model settings. The checkpoints for reference repo can be downloaded here.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train TANet model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/tanet/tanet_imagenet-pretrained-r50_8xb8-dense-1x1x8-100e_kinetics400-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test TANet model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/tanet/tanet_imagenet-pretrained-r50_8xb8-dense-1x1x8-100e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@article{liu2020tam,
  title={TAM: Temporal Adaptive Module for Video Recognition},
  author={Liu, Zhaoyang and Wang, Limin and Wu, Wayne and Qian, Chen and Lu, Tong},
  journal={arXiv preprint arXiv:2005.06803},
  year={2020}
}

TimeSformer¶

Is Space-Time Attention All You Need for Video Understanding?

Abstract¶

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named “TimeSformer,” adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that “divided attention,” where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long).

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8x32x1	224x224	8	TimeSformer (divST)	ImageNet-21K	77.69	93.45	1 clip x 3 crop	196G	122M	config	ckpt	log
8x32x1	224x224	8	TimeSformer (jointST)	ImageNet-21K	76.95	93.28	1 clip x 3 crop	180G	86.11M	config	ckpt	log
8x32x1	224x224	8	TimeSformer (spaceOnly)	ImageNet-21K	76.93	92.88	1 clip x 3 crop	141G	86.11M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
We keep the test setting with the original repo (three crop x 1 clip).
The pretrained model vit_base_patch16_224.pth used by TimeSformer was converted from vision_transformer.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train TimeSformer model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/timesformer/timesformer_divST_8xb8-8x32x1-15e_kinetics400-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test TimeSformer model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/timesformer/timesformer_divST_8xb8-8x32x1-15e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@misc{bertasius2021spacetime,
    title   = {Is Space-Time Attention All You Need for Video Understanding?},
    author  = {Gedas Bertasius and Heng Wang and Lorenzo Torresani},
    year    = {2021},
    eprint  = {2102.05095},
    archivePrefix = {arXiv},
    primaryClass = {cs.CV}
}

TIN¶

Temporal Interlacing Network

Abstract¶

For a long time, the vision community tries to learn the spatio-temporal representation by combining convolutional neural network together with various temporal models, such as the families of Markov chain, optical flow, RNN and temporal convolution. However, these pipelines consume enormous computing resources due to the alternately learning process for spatial and temporal information. One natural question is whether we can embed the temporal information into the spatial one so the information in the two domains can be jointly learned once-only. In this work, we answer this question by presenting a simple yet powerful operator – temporal interlacing network (TIN). Instead of learning the temporal features, TIN fuses the two kinds of information by interlacing spatial representations from the past to the future, and vice versa. A differentiable interlacing target can be learned to control the interlacing process. In this way, a heavy temporal model is replaced by a simple interlacing operator. We theoretically prove that with a learnable interlacing target, TIN performs equivalently to the regularized temporal convolution network (r-TCN), but gains 4% more accuracy with 6x less latency on 6 challenging benchmarks. These results push the state-of-the-art performances of video understanding by a considerable margin. Not surprising, the ensemble model of the proposed TIN won the 1st place in the ICCV19 - Multi Moments in Time challenge.

Results and Models¶

Something-Something V1¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	inference time(video/s)	gpu_mem(M)	config	ckpt	log
1x1x8	height 100	8x4	ResNet50	ImageNet	39.68	68.55	44.04	72.72	8 clips x 1 crop	x	6181	config	ckpt	log

Something-Something V2¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	inference time(video/s)	gpu_mem(M)	config	ckpt	log
1x1x8	height 240	8x4	ResNet50	ImageNet	54.78	82.18	56.48	83.45	8 clips x 1 crop	x	6185	config	ckpt	log

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	inference time(video/s)	gpu_mem(M)	config	ckpt	log
1x1x8	short-side 256	8x4	ResNet50	TSM-Kinetics400	71.86	90.44	8 clips x 1 crop	x	6185	config	ckpt	log

Here, we use finetune to indicate that we use TSM model trained on Kinetics-400 to finetune the TIN model on Kinetics-400.

备注

The reference topk acc are got by training the original repo #1aacd0c with no AverageMeter issue. The AverageMeter issue will lead to incorrect performance, so we fix it before running.
The gpus indicates the number of gpu we used to get the checkpoint. It is noteworthy that the configs we provide are used for 8 gpus as default. According to the Linear Scaling Rule, you may set the learning rate proportional to the batch size if you use different GPUs or videos per GPU, e.g., lr=0.01 for 4 GPUs x 2 video/gpu and lr=0.08 for 16 GPUs x 4 video/gpu.
The values in columns named after “reference” are the results got by training on the original repo, using the same model settings.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train TIN model on Something-Something V1 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/tin/tin_imagenet-pretrained-r50_8xb6-1x1x8-40e_sthv1-rgb.py \
    --work-dir work_dirs/tin_imagenet-pretrained-r50_8xb6-1x1x8-40e_sthv1-rgb randomness.seed=0 randomness.deterministic=True

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test TIN model on Something-Something V1 dataset and dump the result to a json file.

python tools/test.py configs/recognition/tin/tin_imagenet-pretrained-r50_8xb6-1x1x8-40e_sthv1-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.json

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@article{shao2020temporal,
    title={Temporal Interlacing Network},
    author={Hao Shao and Shengju Qian and Yu Liu},
    year={2020},
    journal={AAAI},
}

TPN¶

Temporal Pyramid Network for Action Recognition

Abstract¶

Visual tempo characterizes the dynamics and the temporal scale of an action. Modeling such visual tempos of different actions facilitates their recognition. Previous works often capture the visual tempo through sampling raw videos at multiple rates and constructing an input-level frame pyramid, which usually requires a costly multi-branch network to handle. In this work we propose a generic Temporal Pyramid Network (TPN) at the feature-level, which can be flexibly integrated into 2D or 3D backbone networks in a plug-and-play manner. Two essential components of TPN, the source of features and the fusion of features, form a feature hierarchy for the backbone so that it can capture action instances at various tempos. TPN also shows consistent improvements over other challenging baselines on several action recognition datasets. Specifically, when equipped with TPN, the 3D ResNet-50 with dense sampling obtains a 2% gain on the validation set of Kinetics-400. A further analysis also reveals that TPN gains most of its improvements on action classes that have large variances in their visual tempos, validating the effectiveness of TPN.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	inference time(video/s)	gpu_mem(M)	config	ckpt	log
8x8x1	short-side 320	8x2	ResNet50	None	74.20	91.48	x	x	10 clips x 3 crop	x	6916	config	ckpt	log
8x8x1	short-side 320	8	ResNet50	ImageNet	76.74	92.57	75.49	92.05	10 clips x 3 crop	x	6916	config	ckpt	log

Something-Something V1¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	inference time(video/s)	gpu_mem(M)	config	ckpt	log
1x1x8	height 100	8x6	ResNet50	TSM	51.87	79.67	x	x	8 clips x 3 crop	x	8828	config	ckpt	log

备注

The gpus indicates the number of gpu we used to get the checkpoint. It is noteworthy that the configs we provide are used for 8 gpus as default. According to the Linear Scaling Rule, you may set the learning rate proportional to the batch size if you use different GPUs or videos per GPU, e.g., lr=0.01 for 4 GPUs x 2 video/gpu and lr=0.08 for 16 GPUs x 4 video/gpu.
The values in columns named after “reference” are the results got by testing the checkpoint released on the original repo and codes, using the same dataset with ours.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train TPN model on Kinetics-400 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/tpn/tpn-slowonly_r50_8xb8-8x8x1-150e_kinetics400-rgb.py \
    --work-dir work_dirs/tpn-slowonly_r50_8xb8-8x8x1-150e_kinetics400-rgb [--validate --seed 0 --deterministic]

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test TPN model on Kinetics-400 dataset and dump the result to a json file.

python tools/test.py configs/recognition/tpn/tpn-slowonly_r50_8xb8-8x8x1-150e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{yang2020tpn,
  title={Temporal Pyramid Network for Action Recognition},
  author={Yang, Ceyuan and Xu, Yinghao and Shi, Jianping and Dai, Bo and Zhou, Bolei},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2020},
}

TRN¶

Temporal Relational Reasoning in Videos

Abstract¶

Temporal relational reasoning, the ability to link meaningful transformations of objects or entities over time, is a fundamental property of intelligent species. In this paper, we introduce an effective and interpretable network module, the Temporal Relation Network (TRN), designed to learn and reason about temporal dependencies between video frames at multiple time scales. We evaluate TRN-equipped networks on activity recognition tasks using three recent video datasets - Something-Something, Jester, and Charades - which fundamentally depend on temporal relational reasoning. Our results demonstrate that the proposed TRN gives convolutional neural networks a remarkable capacity to discover temporal relations in videos. Through only sparsely sampled video frames, TRN-equipped networks can accurately predict human-object interactions in the Something-Something dataset and identify various human gestures on the Jester dataset with very competitive performance. TRN-equipped networks also outperform two-stream networks and 3D convolution networks in recognizing daily activities in the Charades dataset. Further analyses show that the models learn intuitive and interpretable visual common sense knowledge in videos.

Results and Models¶

Something-Something V1¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc (efficient/accurate)	top5 acc (efficient/accurate)	testing protocol	FLOPs	params	config	ckpt	log
1x1x8	224x224	8	ResNet50	ImageNet	31.60 / 33.65	60.15 / 62.22	16 clips x 10 crop	42.94G	26.64M	config	ckpt	log

Something-Something V2¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc (efficient/accurate)	top5 acc (efficient/accurate)	testing protocol	FLOPs	params	config	ckpt	log
1x1x8	224x224	8	ResNet50	ImageNet	47.65 / 51.20	76.27 / 78.42	16 clips x 10 crop	42.94G	26.64M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
There are two kinds of test settings for Something-Something dataset, efficient setting (center crop only) and accurate setting (three crop and twice_sample).
In the original repository, the author augments data with random flipping on something-something dataset, but the augmentation method may be wrong due to the direct actions, such as push left to right. So, we replaced flip with flip with label mapping, and change the testing method TenCrop, which has five flipped crops, to Twice Sample & ThreeCrop.
We use ResNet50 instead of BNInception as the backbone of TRN. When Training TRN-ResNet50 on sthv1 dataset in the original repository, we get top1 (top5) accuracy 30.542 (58.627) vs. ours 31.81 (60.47).

For more details on data preparation, you can refer to Something-something V1 and Something-something V2.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train TRN model on sthv1 dataset in a deterministic option with periodic validation.

python tools/train.py configs/recognition/trn/trn_imagenet-pretrained-r50_8xb16-1x1x8-50e_sthv1-rgb.py \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test TRN model on sthv1 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/trn/trn_imagenet-pretrained-r50_8xb16-1x1x8-50e_sthv1-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@article{zhou2017temporalrelation,
    title = {Temporal Relational Reasoning in Videos},
    author = {Zhou, Bolei and Andonian, Alex and Oliva, Aude and Torralba, Antonio},
    journal={European Conference on Computer Vision},
    year={2018}
}

TSM¶

TSM: Temporal Shift Module for Efficient Video Understanding

Abstract¶

The explosive growth in video streaming gives rise to challenges on performing video understanding at high accuracy and low computation cost. Conventional 2D CNNs are computationally cheap but cannot capture temporal relationships; 3D CNN based methods can achieve good performance but are computationally intensive, making it expensive to deploy. In this paper, we propose a generic and effective Temporal Shift Module (TSM) that enjoys both high efficiency and high performance. Specifically, it can achieve the performance of 3D CNN but maintain 2D CNN’s complexity. TSM shifts part of the channels along the temporal dimension; thus facilitate information exchanged among neighboring frames. It can be inserted into 2D CNNs to achieve temporal modeling at zero computation and zero parameters. We also extended TSM to online setting, which enables real-time low-latency online video recognition and video object detection. TSM is accurate and efficient: it ranks the first place on the Something-Something leaderboard upon publication; on Jetson Nano and Galaxy Note8, it achieves a low latency of 13ms and 35ms for online video recognition.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
1x1x8	224x224	8	ResNet50	ImageNet	73.18	90.56	8 clips x 10 crop	32.88G	23.87M	config	ckpt	log
1x1x8	224x224	8	ResNet50	ImageNet	73.22	90.22	8 clips x 10 crop	32.88G	23.87M	config	ckpt	log
1x1x16	224x224	8	ResNet50	ImageNet	75.12	91.55	16 clips x 10 crop	65.75G	23.87M	config	ckpt	log
1x1x8 (dense)	224x224	8	ResNet50	ImageNet	73.38	90.78	8 clips x 10 crop	32.88G	23.87M	config	ckpt	log
1x1x8	224x224	8	ResNet50 (NonLocalDotProduct)	ImageNet	74.49	91.15	8 clips x 10 crop	61.30G	31.68M	config	ckpt	log
1x1x8	224x224	8	ResNet50 (NonLocalGauss)	ImageNet	73.66	90.99	8 clips x 10 crop	59.06G	28.00M	config	ckpt	log
1x1x8	224x224	8	ResNet50 (NonLocalEmbedGauss)	ImageNet	74.34	91.23	8 clips x 10 crop	61.30G	31.68M	config	ckpt	log
1x1x8	224x224	8	MobileNetV2	ImageNet	68.71	88.32	8 clips x 3 crop	3.269G	2.736M	config	ckpt	log
1x1x16	224x224	8	MobileOne-S4	ImageNet	74.38	91.71	16 clips x 10 crop	48.65G	13.72M	config	ckpt	log

Something-something V2¶

frame sampling strategy	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
1x1x8	224x224	8	ResNet50	ImageNet	62.72	87.70	8 clips x 3 crop	32.88G	23.87M	config	ckpt	log
1x1x16	224x224	8	ResNet50	ImageNet	64.16	88.61	16 clips x 3 crop	65.75G	23.87M	config	ckpt	log
1x1x8	224x224	8	ResNet101	ImageNet	63.70	88.28	8 clips x 3 crop	62.66G	42.86M	config	ckpt	log

The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.
MoibleOne backbone supports reparameterization during inference. You can use the provided reparameterize tool to convert the checkpoint and switch to the deploy config file.

For more details on data preparation, you can refer to Kinetics400.

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train TSM model on Kinetics-400 dataset in a deterministic option.

python tools/train.py configs/recognition/tsm/tsm_imagenet-pretrained-r50_8xb16-1x1x8-50e_kinetics400-rgb.py \
     --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test TSM model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/tsm/tsm_imagenet-pretrained-r50_8xb16-1x1x8-50e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{lin2019tsm,
  title={TSM: Temporal Shift Module for Efficient Video Understanding},
  author={Lin, Ji and Gan, Chuang and Han, Song},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  year={2019}
}

@article{Nonlocal2018,
  author =   {Xiaolong Wang and Ross Girshick and Abhinav Gupta and Kaiming He},
  title =    {Non-local Neural Networks},
  journal =  {CVPR},
  year =     {2018}
}

TSN¶

Temporal segment networks: Towards good practices for deep action recognition

Abstract¶

Deep convolutional networks have achieved great success for visual recognition in still images. However, for action recognition in videos, the advantage over traditional methods is not so evident. This paper aims to discover the principles to design effective ConvNet architectures for action recognition in videos and learn these models given limited training samples. Our first contribution is temporal segment network (TSN), a novel framework for video-based action recognition. which is based on the idea of long-range temporal structure modeling. It combines a sparse temporal sampling strategy and video-level supervision to enable efficient and effective learning using the whole action video. The other contribution is our study on a series of good practices in learning ConvNets on video data with the help of temporal segment network. Our approach obtains the state-the-of-art performance on the datasets of HMDB51 ( 69.4%) and UCF101 (94.2%). We also visualize the learned ConvNet models, which qualitatively demonstrates the effectiveness of temporal segment network and the proposed good practices.

Results and Models¶

Kinetics-400¶

frame sampling strategy	scheduler	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
1x1x3	MultiStep	224x224	8	ResNet50	ImageNet	72.83	90.65	25 clips x 10 crop	102.7G	24.33M	config	ckpt	log
1x1x5	MultiStep	224x224	8	ResNet50	ImageNet	73.80	91.21	25 clips x 10 crop	102.7G	24.33M	config	ckpt	log
1x1x8	MultiStep	224x224	8	ResNet50	ImageNet	74.12	91.34	25 clips x 10 crop	102.7G	24.33M	config	ckpt	log
dense-1x1x5	MultiStep	224x224	8	ResNet50	ImageNet	71.37	89.67	25 clips x 10 crop	102.7G	24.33M	config	ckpt	log
1x1x8	MultiStep	224x224	8	ResNet101	ImageNet	75.89	92.07	25 clips x 10 crop	195.8G	43.32M	config	ckpt	log

Something-Something V2¶

frame sampling strategy	scheduler	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
1x1x8	MultiStep	224x224	8	ResNet50	ImageNet	35.51	67.09	25 clips x 10 crop	102.7G	24.33M	config	ckpt	log
1x1x16	MultiStep	224x224	8	ResNet50	ImageNet	36.91	68.77	25 clips x 10 crop	102.7G	24.33M	config	ckpt	log

Using backbones from 3rd-party in TSN¶

It’s possible and convenient to use a 3rd-party backbone for TSN under the framework of MMAction2, here we provide some examples for:

[x] Backbones from MMClassification
[x] Backbones from MMPretrain
[x] Backbones from TorchVision
[x] Backbones from TIMM (pytorch-image-models)

frame sampling strategy	scheduler	resolution	gpus	backbone	pretrain	top1 acc	top5 acc	testing protocol	FLOPs	params	config	ckpt	log
1x1x3	MultiStep	224x224	8	ResNext101	ImageNet	72.95	90.36	25 clips x 10 crop	200.3G	42.95M	config	ckpt	log
1x1x3	MultiStep	224x224	8	DenseNet161	ImageNet	72.07	90.15	25 clips x 10 crop	194.6G	27.36M	config	ckpt	log
1x1x3	MultiStep	224x224	8	Swin Transformer	ImageNet	77.03	92.61	25 clips x 10 crop	386.7G	87.15M	config	ckpt	log
1x1x8	MultiStep	224x224	8	Swin Transformer	ImageNet	79.22	94.20	25 clips x 10 crop	386.7G	87.15M	config	ckpt	log
1x1x8	MultiStep	224x224	8	MobileOne-S4	ImageNet	73.65	91.32	25 clips x 10 crop	76G	13.72M	config	ckpt	log

Note that some backbones in TIMM are not supported due to multiple reasons. Please refer to PR #880 for details.
The gpus indicates the number of gpus we used to get the checkpoint. If you want to use a different number of gpus or videos per gpu, the best way is to set --auto-scale-lr when calling tools/train.py, this parameter will auto-scale the learning rate according to the actual batch size and the original batch size.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.
MoibleOne backbone supports reparameterization during inference. You can use the provided reparameterize tool to convert the checkpoint and switch to the deploy config file.

For more details on data preparation, you can refer to

Train¶

You can use the following command to train a model.

python tools/train.py ${CONFIG_FILE} [optional arguments]

Example: train TSN model on Kinetics-400 dataset in a deterministic option.

python tools/train.py configs/recognition/tsn/tsn_imagenet-pretrained-r50_8xb32-1x1x3-100e_kinetics400-rgb.py  \
    --seed=0 --deterministic

For more details, you can refer to the Training part in the Training and Test Tutorial.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test TSN model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/tsn/tsn_imagenet-pretrained-r50_8xb32-1x1x3-100e_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{wang2016temporal,
  title={Temporal segment networks: Towards good practices for deep action recognition},
  author={Wang, Limin and Xiong, Yuanjun and Wang, Zhe and Qiao, Yu and Lin, Dahua and Tang, Xiaoou and Van Gool, Luc},
  booktitle={European conference on computer vision},
  pages={20--36},
  year={2016},
  organization={Springer}
}

UniFormer¶

UniFormer: Unified Transformer for Efficient Spatiotemporal Representation Learning

Abstract¶

It is a challenging task to learn rich and multi-scale spatiotemporal semantics from high-dimensional videos, due to large local redundancy and complex global dependency between video frames. The recent advances in this research have been mainly driven by 3D convolutional neural networks and vision transformers. Although 3D convolution can efficiently aggregate local context to suppress local redundancy from a small 3D neighborhood, it lacks the capability to capture global dependency because of the limited receptive field. Alternatively, vision transformers can effectively capture long-range dependency by self-attention mechanism, while having the limitation on reducing local redundancy with blind similarity comparison among all the tokens in each layer. Based on these observations, we propose a novel Unified transFormer (UniFormer) which seamlessly integrates merits of 3D convolution and spatiotemporal self-attention in a concise transformer format, and achieves a preferable balance between computation and accuracy. Different from traditional transformers, our relation aggregator can tackle both spatiotemporal redundancy and dependency, by learning local and global token affinity respectively in shallow and deep layers. We conduct extensive experiments on the popular video benchmarks, e.g., Kinetics-400, Kinetics-600, and Something-Something V1&V2. With only ImageNet-1K pretraining, our UniFormer achieves 82.9%/84.8% top-1 accuracy on Kinetics-400/Kinetics-600, while requiring 10x fewer GFLOPs than other state-of-the-art methods. For Something-Something V1 and V2, our UniFormer achieves new state-of-the-art performances of 60.9% and 71.2% top-1 accuracy respectively.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	backbone	top1 acc	top5 acc	reference top1 acc	reference top5 acc	mm-Kinetics top1 acc	mm-Kinetics top5 acc	testing protocol	FLOPs	params	config	ckpt
16x4x1	short-side 320	UniFormer-S	80.9	94.6	80.8	94.7	80.9	94.6	4 clips x 1 crop	41.8G	21.4M	config	ckpt
16x4x1	short-side 320	UniFormer-B	82.0	95.0	82.0	95.1	82.0	95.0	4 clips x 1 crop	96.7G	49.8M	config	ckpt
32x4x1	short-side 320	UniFormer-B	83.1	95.3	82.9	95.4	83.0	95.3	4 clips x 1 crop	59G	49.8M	config	ckpt

The models are ported from the repo UniFormer and tested on our data. Currently, we only support the testing of UniFormer models, training will be available soon.

The values in columns named after “reference” are the results of the original repo.
The values in top1/5 acc is tested on the same data list as the original repo, and the label map is provided by UniFormer. The total videos are available at Kinetics400 (BaiduYun password: g5kp), which consists of 19787 videos.
The values in columns named after “mm-Kinetics” are the testing results on the Kinetics dataset held by MMAction2, which is also used by other models in MMAction2. Due to the differences between various versions of Kinetics dataset, there is a little gap between top1/5 acc and mm-Kinetics top1/5 acc. For a fair comparison with other models, we report both results here. Note that we simply report the inference results, since the training set is different between UniFormer and other models, the results are lower than that tested on the author’s version.
Since the original models for Kinetics-400/600/700 adopt different label file, we simply map the weight according to the label name. New label map for Kinetics-400/600/700 can be found here.
Due to some difference between SlowFast and MMAction2, there are some gaps between their performances.

For more details on data preparation, you can refer to preparing_kinetics.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test UniFormer-S model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/uniformer/uniformer-small_imagenet1k-pre_16x4x1_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{
  li2022uniformer,
  title={UniFormer: Unified Transformer for Efficient Spatial-Temporal Representation Learning},
  author={Kunchang Li and Yali Wang and Gao Peng and Guanglu Song and Yu Liu and Hongsheng Li and Yu Qiao},
  booktitle={International Conference on Learning Representations},
  year={2022},
  url={https://openreview.net/forum?id=nBU_u6DLvoK}
}

UniFormerV2¶

UniFormerV2: Spatiotemporal Learning by Arming Image ViTs with Video UniFormer

Abstract¶

Learning discriminative spatiotemporal representation is the key problem of video understanding. Recently, Vision Transformers (ViTs) have shown their power in learning long-term video dependency with self-attention. Unfortunately, they exhibit limitations in tackling local video redundancy, due to the blind global comparison among tokens. UniFormer has successfully alleviated this issue, by unifying convolution and self-attention as a relation aggregator in the transformer format. However, this model has to require a tiresome and complicated image-pretraining phrase, before being finetuned on videos. This blocks its wide usage in practice. On the contrary, open-sourced ViTs are readily available and well-pretrained with rich image supervision. Based on these observations, we propose a generic paradigm to build a powerful family of video networks, by arming the pretrained ViTs with efficient UniFormer designs. We call this family UniFormerV2, since it inherits the concise style of the UniFormer block. But it contains brand-new local and global relation aggregators, which allow for preferable accuracy-computation balance by seamlessly integrating advantages from both ViTs and UniFormer. Without any bells and whistles, our UniFormerV2 gets the state-of-the-art recognition performance on 8 popular video benchmarks, including scene-related Kinetics-400/600/700 and Moments in Time, temporal-related Something-Something V1/V2, untrimmed ActivityNet and HACS. In particular, it is the first model to achieve 90% top-1 accuracy on Kinetics-400, to our best knowledge.

Results and Models¶

Kinetics-400¶

uniform sampling	resolution	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	mm-Kinetics top1 acc	mm-Kinetics top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8	short-side 320	UniFormerV2-B/16	clip	-	-	84.3	96.4	84.4	96.3	4 clips x 3 crop	0.1T	115M	config	ckpt	log
8	short-side 320	UniFormerV2-B/16	clip-kinetics710	-	-	85.6	97.0	85.8	97.1	4 clips x 3 crop	0.1T	115M	config	ckpt	log
8	short-side 320	UniFormerV2-L/14*	clip-kinetics710	88.7	98.1	88.8	98.1	88.7	98.1	4 clips x 3 crop	0.7T	354M	config	ckpt	-
16	short-side 320	UniFormerV2-L/14*	clip-kinetics710	89.0	98.2	89.1	98.2	89.0	98.2	4 clips x 3 crop	1.3T	354M	config	ckpt	-
32	short-side 320	UniFormerV2-L/14*	clip-kinetics710	89.3	98.2	89.3	98.2	89.4	98.2	2 clips x 3 crop	2.7T	354M	config	ckpt	-
32	short-side 320	UniFormerV2-L/14@336*	clip-kinetics710	89.5	98.4	89.7	98.3	89.5	98.4	2 clips x 3 crop	6.3T	354M	config	ckpt	-

Kinetics-600¶

uniform sampling	resolution	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	mm-Kinetics top1 acc	mm-Kinetics top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8	Raw	UniFormerV2-B/16	clip-kinetics710	-	-	86.1	97.2	86.4	97.3	4 clips x 3 crop	0.1T	115M	config	ckpt	log
8	Raw	UniFormerV2-L/14*	clip-kinetics710	89.0	98.3	89.0	98.2	87.5	98.0	4 clips x 3 crop	0.7T	354M	config	ckpt	-
16	Raw	UniFormerV2-L/14*	clip-kinetics710	89.4	98.3	89.4	98.3	87.8	98.0	4 clips x 3 crop	1.3T	354M	config	ckpt	-
32	Raw	UniFormerV2-L/14*	clip-kinetics710	89.2	98.3	89.5	98.3	87.7	98.1	2 clips x 3 crop	2.7T	354M	config	ckpt	-
32	Raw	UniFormerV2-L/14@336*	clip-kinetics710	89.8	98.5	89.9	98.5	88.8	98.3	2 clips x 3 crop	6.3T	354M	config	ckpt	-

Kinetics-700¶

uniform sampling	resolution	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	mm-Kinetics top1 acc	mm-Kinetics top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8	Raw	UniFormerV2-B/16	clip	-	-	75.8	92.8	75.9	92.9	4 clips x 3 crop	0.1T	115M	config	ckpt	log
8	Raw	UniFormerV2-B/16	clip-kinetics710	-	-	76.3	92.7	76.3	92.9	4 clips x 3 crop	0.1T	115M	config	ckpt	log
8	Raw	UniFormerV2-L/14*	clip-kinetics710	80.8	95.2	80.8	95.4	79.4	94.8	4 clips x 3 crop	0.7T	354M	config	ckpt	-
16	Raw	UniFormerV2-L/14*	clip-kinetics710	81.2	95.6	81.2	95.6	79.2	95.0	4 clips x 3 crop	1.3T	354M	config	ckpt	-
32	Raw	UniFormerV2-L/14*	clip-kinetics710	81.4	95.7	81.5	95.7	79.8	95.3	2 clips x 3 crop	2.7T	354M	config	ckpt	-
32	Raw	UniFormerV2-L/14@336*	clip-kinetics710	82.1	96.0	82.1	96.1	80.6	95.6	2 clips x 3 crop	6.3T	354M	config	ckpt	-

MiTv1¶

uniform sampling	resolution	backbone	pretrain	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	FLOPs	params	config	ckpt	log
8	Raw	UniFormerV2-B/16	clip-kinetics710-kinetics400	42.3	71.5	42.6	71.7	4 clips x 3 crop	0.1T	115M	config	ckpt	log
8	Raw	UniFormerV2-L/14*	clip-kinetics710-kinetics400	47.0	76.1	47.0	76.1	4 clips x 3 crop	0.7T	354M	config	ckpt	-
8	Raw	UniFormerV2-L/14@336*	clip-kinetics710-kinetics400	47.7	76.8	47.8	76.0	4 clips x 3 crop	1.6T	354M	config	ckpt	-

Kinetics-710¶

uniform sampling	resolution	backbone	pretrain	top1 acc	top5 acc	config	ckpt	log
8	Raw	UniFormerV2-B/16*	clip	78.9	94.2	config	ckpt	log
8	Raw	UniFormerV2-L/14*	clip	-	-	config	ckpt	-
8	Raw	UniFormerV2-L/14@336*	clip	-	-	config	ckpt	-

The models with * are ported from the repo UniFormerV2 and tested on our data. Due to computational limitations, we only support reliable training config for base model (i.e. UniFormerV2-B/16).

The values in columns named after “reference” are the results of the original repo.
The values in top1/5 acc is tested on the same data list as the original repo, and the label map is provided by UniFormerV2.
The values in columns named after “mm-Kinetics” are the testing results on the Kinetics dataset held by MMAction2, which is also used by other models in MMAction2. Due to the differences between various versions of Kinetics dataset, there is a little gap between top1/5 acc and mm-Kinetics top1/5 acc. For a fair comparison with other models, we report both results here. Note that we simply report the inference results, since the training set is different between UniFormer and other models, the results are lower than that tested on the author’s version.
Since the original models for Kinetics-400/600/700 adopt different label file, we simply map the weight according to the label name. New label map for Kinetics-400/600/700 can be found here.
Due to some differences between SlowFast and MMAction2, there are some gaps between their performances.
Kinetics-710 is used for pretraining, which helps improve the performance on other datasets efficiently. You can find more details in the paper. We also map the wegiht for Kinetics-710 checkpoints, you can find the label map here.

For more details on data preparation, you can refer to

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test UniFormerV2-B/16 model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/uniformerv2/uniformerv2-base-p16-res224_clip-kinetics710-pre_u8_kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@article{Li2022UniFormerV2SL,
  title={UniFormerV2: Spatiotemporal Learning by Arming Image ViTs with Video UniFormer},
  author={Kunchang Li and Yali Wang and Yinan He and Yizhuo Li and Yi Wang and Limin Wang and Y. Qiao},
  journal={ArXiv},
  year={2022},
  volume={abs/2211.09552}
}

VideoMAE¶

VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training

Abstract¶

Pre-training video transformers on extra large-scale datasets is generally required to achieve premier performance on relatively small datasets. In this paper, we show that video masked autoencoders (VideoMAE) are data-efficient learners for self-supervised video pre-training (SSVP). We are inspired by the recent ImageMAE and propose customized video tube masking with an extremely high ratio. This simple design makes video reconstruction a more challenging self-supervision task, thus encouraging extracting more effective video representations during this pre-training process. We obtain three important findings on SSVP: (1) An extremely high proportion of masking ratio (i.e., 90% to 95%) still yields favorable performance of VideoMAE. The temporally redundant video content enables a higher masking ratio than that of images. (2) VideoMAE achieves impressive results on very small datasets (i.e., around 3k-4k videos) without using any extra data. (3) VideoMAE shows that data quality is more important than data quantity for SSVP. Domain shift between pre-training and target datasets is an important issue. Notably, our VideoMAE with the vanilla ViT can achieve 87.4% on Kinetics-400, 75.4% on Something-Something V2, 91.3% on UCF101, and 62.6% on HMDB51, without using any extra data.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	backbone	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	FLOPs	params	config	ckpt
16x4x1	short-side 320	ViT-B	81.3	95.0	81.5 [VideoMAE]	95.1 [VideoMAE]	5 clips x 3 crops	180G	87M	config	ckpt [1]
16x4x1	short-side 320	ViT-L	85.3	96.7	85.2 [VideoMAE]	96.8 [VideoMAE]	5 clips x 3 crops	597G	305M	config	ckpt [1]

[1] The models are ported from the repo VideoMAE and tested on our data. Currently, we only support the testing of VideoMAE models, training will be available soon.

The values in columns named after “reference” are the results of the original repo.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to preparing_kinetics.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test ViT-base model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/videomae/vit-base-p16_videomae-k400-pre_16x4x1_kinetics-400.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@inproceedings{tong2022videomae,
  title={Video{MAE}: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training},
  author={Zhan Tong and Yibing Song and Jue Wang and Limin Wang},
  booktitle={Advances in Neural Information Processing Systems},
  year={2022}
}

VideoMAE V2¶

VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking

Abstract¶

Scale is the primary factor for building a powerful foundation model that could well generalize to a variety of downstream tasks. However, it is still challenging to train video foundation models with billions of parameters. This paper shows that video masked autoencoder (VideoMAE) is a scalable and general self-supervised pre-trainer for building video foundation models. We scale the VideoMAE in both model and data with a core design. Specifically, we present a dual masking strategy for efficient pre-training, with an encoder operating on a subset of video tokens and a decoder processing another subset of video tokens. Although VideoMAE is very efficient due to high masking ratio in encoder, masking decoder can still further reduce the overall computational cost. This enables the efficient pre-training of billion-level models in video. We also use a progressive training paradigm that involves an initial pre-training on a diverse multi-sourced unlabeled dataset, followed by a post-pre-training on a mixed labeled dataset. Finally, we successfully train a video ViT model with a billion parameters, which achieves a new state-of-the-art performance on the datasets of Kinetics (90.0% on K400 and 89.9% on K600) and Something-Something (68.7% on V1 and 77.0% on V2). In addition, we extensively verify the pre-trained video ViT models on a variety of downstream tasks, demonstrating its effectiveness as a general video representation learner.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	backbone	top1 acc	top5 acc	reference top1 acc	reference top5 acc	testing protocol	FLOPs	params	config	ckpt
16x4x1	short-side 320	ViT-S	83.6	96.3	83.7 [VideoMAE V2]	96.2 [VideoMAE V2]	5 clips x 3 crops	57G	22M	config	ckpt [1]
16x4x1	short-side 320	ViT-B	86.6	97.3	86.6 [VideoMAE V2]	97.3 [VideoMAE V2]	5 clips x 3 crops	180G	87M	config	ckpt [1]

[1] The models were distilled from the VideoMAE V2-g model. Specifically, models are initialized with VideoMAE V2 pretraining, then distilled on Kinetics 710 dataset. They are ported from the repo VideoMAE V2 and tested on our data. The VideoMAE V2-g model can be obtained from the original repository. Currently, we only support the testing of VideoMAE V2 models.

The values in columns named after “reference” are the results of the original repo.
The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at Kinetics400-Validation. The corresponding data list (each line is of the format ‘video_id, num_frames, label_index’) and the label map are also available.

For more details on data preparation, you can refer to preparing_kinetics.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test ViT-base model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/videomaev2/vit-base-p16_videomaev2-vit-g-dist-k710-pre_16x4x1_kinetics-400.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@misc{wang2023videomaev2,
      title={VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking},
      author={Limin Wang and Bingkun Huang and Zhiyu Zhao and Zhan Tong and Yinan He and Yi Wang and Yali Wang and Yu Qiao},
      year={2023},
      eprint={2303.16727},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

X3D¶

X3D: Expanding Architectures for Efficient Video Recognition

Abstract¶

This paper presents X3D, a family of efficient video networks that progressively expand a tiny 2D image classification architecture along multiple network axes, in space, time, width and depth. Inspired by feature selection methods in machine learning, a simple stepwise network expansion approach is employed that expands a single axis in each step, such that good accuracy to complexity trade-off is achieved. To expand X3D to a specific target complexity, we perform progressive forward expansion followed by backward contraction. X3D achieves state-of-the-art performance while requiring 4.8x and 5.5x fewer multiply-adds and parameters for similar accuracy as previous work. Our most surprising finding is that networks with high spatiotemporal resolution can perform well, while being extremely light in terms of network width and parameters. We report competitive accuracy at unprecedented efficiency on video classification and detection benchmarks.

Results and Models¶

Kinetics-400¶

frame sampling strategy	resolution	backbone	top1 10-view	top1 30-view	reference top1 10-view	reference top1 30-view	config	ckpt
13x6x1	160x160	X3D_S	73.2	73.3	73.1 [SlowFast]	73.5 [SlowFast]	config	ckpt[1]
16x5x1	224x224	X3D_M	75.2	76.4	75.1 [SlowFast]	76.2 [SlowFast]	config	ckpt[1]

[1] The models are ported from the repo SlowFast and tested on our data. Currently, we only support the testing of X3D models, training will be available soon.

The values in columns named after “reference” are the results got by testing the checkpoint released on the original repo and codes, using the same dataset with ours.
The validation set of Kinetics400 we used is same as the repo SlowFast, which is available here.

For more details on data preparation, you can refer to Kinetics400.

Test¶

You can use the following command to test a model.

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [optional arguments]

Example: test X3D model on Kinetics-400 dataset and dump the result to a pkl file.

python tools/test.py configs/recognition/x3d/x3d_s_13x6x1_facebook-kinetics400-rgb.py \
    checkpoints/SOME_CHECKPOINT.pth --dump result.pkl

For more details, you can refer to the Test part in the Training and Test Tutorial.

Citation¶

@misc{feichtenhofer2020x3d,
      title={X3D: Expanding Architectures for Efficient Video Recognition},
      author={Christoph Feichtenhofer},
      year={2020},
      eprint={2004.04730},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}