Is Space-Time Attention All You Need for Video Understanding?
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).
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 callingtools/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.
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.
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.
@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}
}