[Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095)
<!-- [ALGORITHM] -->
## Abstract
<!-- [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).
1. The **gpus** indicates the number of gpu (32G V100) 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](https://arxiv.org/abs/1706.02677), you may set the learning rate proportional to the batch size if you use different GPUs or videos per GPU,
e.g., lr=0.005 for 8 GPUs x 8 videos/gpu and lr=0.00375 for 8 GPUs x 6 videos/gpu.
2. We keep the test setting with the [original repo](https://github.com/facebookresearch/TimeSformer)(three crop x 1 clip).
3. The pretrained model `vit_base_patch16_224.pth` used by TimeSformer was converted from [vision_transformer](https://github.com/google-research/vision_transformer).
:::
For more details on data preparation, you can refer to Kinetics400 in [Data Preparation](/docs/data_preparation.md).
## Train
You can use the following command to train a model.
Training Json Log:https://download.openmmlab.com/mmaction/recognition/timesformer/timesformer_divST_8x32x1_15e_kinetics400_rgb/timesformer_divST_8x32x1_15e_kinetics400_rgb.json
Training Log:https://download.openmmlab.com/mmaction/recognition/timesformer/timesformer_divST_8x32x1_15e_kinetics400_rgb/timesformer_divST_8x32x1_15e_kinetics400_rgb.log
Training Json Log:https://download.openmmlab.com/mmaction/recognition/timesformer/timesformer_jointST_8x32x1_15e_kinetics400_rgb/timesformer_jointST_8x32x1_15e_kinetics400_rgb.json
Training Log:https://download.openmmlab.com/mmaction/recognition/timesformer/timesformer_jointST_8x32x1_15e_kinetics400_rgb/timesformer_jointST_8x32x1_15e_kinetics400_rgb.log
Training Json Log:https://download.openmmlab.com/mmaction/recognition/timesformer/timesformer_spaceOnly_8x32x1_15e_kinetics400_rgb/timesformer_spaceOnly_8x32x1_15e_kinetics400_rgb.json
Training Log:https://download.openmmlab.com/mmaction/recognition/timesformer/timesformer_spaceOnly_8x32x1_15e_kinetics400_rgb/timesformer_spaceOnly_8x32x1_15e_kinetics400_rgb.log
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.
Here, we use `finetune` to indicate that we use [TSM model](https://download.openmmlab.com/mmaction/recognition/tsm/tsm_r50_1x1x8_50e_kinetics400_rgb/tsm_r50_1x1x8_50e_kinetics400_rgb_20200607-af7fb746.pth) trained on Kinetics-400 to finetune the TIN model on Kinetics-400.
:::{note}
1. The **reference topk acc** are got by training the [original repo #1aacd0c](https://github.com/deepcs233/TIN/tree/1aacd0c4c30d5e1d334bf023e55b855b59f158db) with no [AverageMeter issue](https://github.com/deepcs233/TIN/issues/4).
The [AverageMeter issue](https://github.com/deepcs233/TIN/issues/4) will lead to incorrect performance, so we fix it before running.
2. 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](https://arxiv.org/abs/1706.02677), 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.
3. The **inference_time** is got by this [benchmark script](/tools/analysis/benchmark.py), where we use the sampling frames strategy of the test setting and only care about the model inference time,
not including the IO time and pre-processing time. For each setting, we use 1 gpu and set batch size (videos per gpu) to 1 to calculate the inference time.
4. The values in columns named after "reference" are the results got by training on the original repo, using the same model settings.
5. The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at [Kinetics400-Validation](https://mycuhk-my.sharepoint.com/:u:/g/personal/1155136485_link_cuhk_edu_hk/EbXw2WX94J1Hunyt3MWNDJUBz-nHvQYhO9pvKqm6g39PMA?e=a9QldB). The corresponding [data list](https://download.openmmlab.com/mmaction/dataset/k400_val/kinetics_val_list.txt)(each line is of the format 'video_id, num_frames, label_index') and the [label map](https://download.openmmlab.com/mmaction/dataset/k400_val/kinetics_class2ind.txt) are also available.
:::
For more details on data preparation, you can refer to Kinetics400, Something-Something V1 and Something-Something V2 in [Data Preparation](/docs/data_preparation.md).
## Train
You can use the following command to train a model.
[Temporal Pyramid Network for Action Recognition](https://openaccess.thecvf.com/content_CVPR_2020/html/Yang_Temporal_Pyramid_Network_for_Action_Recognition_CVPR_2020_paper.html)
<!-- [ALGORITHM] -->
## Abstract
<!-- [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.
1. 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](https://arxiv.org/abs/1706.02677), 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.
2. The **inference_time** is got by this [benchmark script](/tools/analysis/benchmark.py), where we use the sampling frames strategy of the test setting and only care about the model inference time,
not including the IO time and pre-processing time. For each setting, we use 1 gpu and set batch size (videos per gpu) to 1 to calculate the inference time.
3. 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.
4. The validation set of Kinetics400 we used consists of 19796 videos. These videos are available at [Kinetics400-Validation](https://mycuhk-my.sharepoint.com/:u:/g/personal/1155136485_link_cuhk_edu_hk/EbXw2WX94J1Hunyt3MWNDJUBz-nHvQYhO9pvKqm6g39PMA?e=a9QldB). The corresponding [data list](https://download.openmmlab.com/mmaction/dataset/k400_val/kinetics_val_list.txt)(each line is of the format 'video_id, num_frames, label_index') and the [label map](https://download.openmmlab.com/mmaction/dataset/k400_val/kinetics_class2ind.txt) are also available.
:::
For more details on data preparation, you can refer to Kinetics400, Something-Something V1 and Something-Something V2 in [Data Preparation](/docs/data_preparation.md).
## Train
You can use the following command to train a model.
Training Json Log:https://download.openmmlab.com/mmaction/recognition/tpn/tpn_slowonly_r50_8x8x1_150e_kinetics_rgb/tpn_slowonly_r50_8x8x1_150e_kinetics_rgb.json
Training Log:https://download.openmmlab.com/mmaction/recognition/tpn/tpn_slowonly_r50_8x8x1_150e_kinetics_rgb/tpn_slowonly_r50_8x8x1_150e_kinetics_rgb.log
Training Json Log:https://download.openmmlab.com/mmaction/recognition/tpn/tpn_imagenet_pretrained_slowonly_r50_8x8x1_150e_kinetics_rgb/tpn_imagenet_pretrained_slowonly_r50_8x8x1_150e_kinetics_rgb.json
Training Log:https://download.openmmlab.com/mmaction/recognition/tpn/tpn_imagenet_pretrained_slowonly_r50_8x8x1_150e_kinetics_rgb/tpn_imagenet_pretrained_slowonly_r50_8x8x1_150e_kinetics_rgb.log
[Temporal Relational Reasoning in Videos](https://openaccess.thecvf.com/content_ECCV_2018/html/Bolei_Zhou_Temporal_Relational_Reasoning_ECCV_2018_paper.html)
<!-- [ALGORITHM] -->
## Abstract
<!-- [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.
1. 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](https://arxiv.org/abs/1706.02677), 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.
2. There are two kinds of test settings for Something-Something dataset, efficient setting (center crop x 1 clip) and accurate setting (Three crop x 2 clip).
3. In the original [repository](https://github.com/zhoubolei/TRN-pytorch), 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`.
4. 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 31.62 (60.01).
:::
For more details on data preparation, you can refer to
-[preparing_sthv1](/tools/data/sthv1/README.md)
-[preparing_sthv2](/tools/data/sthv2/README.md)
## Train
You can use the following command to train a model.
