[ICCV'2023 Oral] Implicit Temporal Modeling with Learnable Alignment for Video Recognition

This is an official implementation of ILA, a new temporal modeling method for video action recognition.

Implicit Temporal Modeling with Learnable Alignment for Video Recognition
accepted by ICCV 2023
Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, Yu-Gang Jiang

[arxiv] [pdf] [supp] [slides]

News

:star2: [August, 2023] ILA has been accepted as ICCV2023 oral presentation.

Environment Setup

To set up the environment, you can easily run the following command:

pip install torch==1.11.0
pip install torchvision==0.12.0
pip install pathlib
pip install mmcv-full
pip install decord
pip install ftfy
pip install einops
pip install termcolor
pip install timm
pip install regex

Install Apex as follows

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

It is worth noting that this branch is for k400. For SSv2, please refer to SSv2 branch.

Data Preparation

For downloading the Kinetics datasets, you can refer to mmaction2 or CVDF. For Something-Something v2, you can get them from the official website.

Due to limited storage, we decord the videos in an online fashion using decord.

We provide the following way to organize the dataset:

Standard Folder: For standard folder, put all videos in the videos folder, and prepare the annotation files as train.txt and val.txt. Please make sure the folder looks like this:
```
$ ls /PATH/TO/videos | head -n 2
a.mp4
b.mp4

$ head -n 2 /PATH/TO/train.txt
a.mp4 0
b.mp4 2

$ head -n 2 /PATH/TO/val.txt
c.mp4 1
d.mp4 2
```

Train

The training configurations lie in configs. For example, you can run the following command to train ILA-ViT-B/16 with 8 frames on Something-Something v2.

python -m torch.distributed.launch --nproc_per_node=8 main.py -cfg configs/ssv2/16_8.yaml --output /PATH/TO/OUTPUT --accumulation-steps 8

Note:

We recommend setting the total batch size to 256.
Please specify the data path in config file(configs/*.yaml). For standard folder, set that to /PATH/TO/videos naturally.
The pretrained CLIP is specified by using --pretrained /PATH/TO/PRETRAINED.

Test

For example, you can run the following command to validate the ILA-ViT-B/16 with 8 frames on Something-Something v2.

python -m torch.distributed.launch --nproc_per_node=8 main.py -cfg configs/ssv2/16_8.yaml --output /PATH/TO/OUTPUT --only_test --resume /PATH/TO/CKPT --opts TEST.NUM_CLIP 4 TEST.NUM_CROP 3

Note:

According to our experience and sanity checks, there is a reasonable random variation about accuracy when testing on different machines.
There are two parts in the provided logs. The first part is conventional training followed by validation per epoch with single-view. The second part refers to the multiview (3 crops x 4 clips) inference logs.

Main Results in paper

This is an original-implementation for open-source use. In the following table we report the accuracy in original paper.

Fully-supervised on Kinetics-400:

Model	Input	Top-1 Acc.(%)	Top-5 Acc.(%)	ckpt
ILA-B/32	8x224	81.3	95.0	GoogleDrive
ILA-B/32	16x224	82.4	95.8	GoogleDrive
ILA-B/16	8x224	84.0	96.6	GoogleDrive
ILA-B/16	16x224	85.7	97.2	GoogleDrive
ILA-L/14	8x224	88.0	98.1	GoogleDrive
ILA-L/14	16x336	88.7	97.8	GoogleDrive

Fully-supervised on Something-Something v2:

Model	Input	Top-1 Acc.(%)	Top-5 Acc.(%)	ckpt
ILA-B/16	8x224	65.0	89.2	GoogleDrive
ILA-B/16	16x224	66.8	90.3	GoogleDrive
ILA-L/14	8x224	67.8	90.5	GoogleDrive
ILA-L/14	16x336	70.2	91.8	GoogleDrive

Bibtex

If this project is useful for you, please consider citing our paper :

@inproceedings{tu2023ila,
  title={Implicit Temporal Modeling with Learnable Alignment for Video Recognition},
  author={Tu, Shuyuan and Dai, Qi and Wu, Zuxuan and Cheng, Zhi-Qi and Hu, Han and Jiang, Yu-Gang},
  booktitle={ICCV},
  year={2023}
}

Acknowledgements

Parts of the codes are borrowed from mmaction2, Swin and X-CLIP. Sincere thanks to their wonderful works.

Francis-Rings / ILA

readme