Awesome

Temporally-Adaptive Convolutions for Efficient Video Understanding

This repository provides the official pytorch implementation of the following papers for video classification and temporal localization. For more details on the respective paper, please refer to the project folder.

Video/Action Classification

• Temporally-Adaptive Models for Efficient Video Understanding, arXiv 2023<br>

• TAda! Temporally-Adaptive Convolutions for Video Understanding, ICLR 2022 [Website]<br>

• Towards Training Stronger Video Vision Transformers for EPIC-KITCHENS-100 Action Recognition<br>CVPRW 2021 Rank 2 submission to EPIC-KITCHENS-100 Action Recognition challenge

Self-supervised video representation learning

• Self-supervised Motion Learning from Static Images, CVPR 2021<br>

• ParamCrop: Parametric Cubic Cropping for Video Contrastive Learning, TMM <br>

Temporal Action Localization

• A Stronger Baseline for Ego-Centric Action Detection<br>CVPRW 2021 Rank 1 submission to EPIC-KITCHENS-100 Action Detection Challenge

Latest

[2023-08] 🔥 Released models for TAdaConvNeXtV2 and TAdaFormer. See MODEL_ZOO and have a try!

[2023-08] Released code for TAdaConvNeXtV2 and TAdaFormer.

[2022-02] TAda2D features for action localization released.

[2022-01] TAdaConv accepted to ICLR 2022.

[2021-10] Codes and models released.

Guidelines

Installation, data preparation and running

The general pipeline for using this repo is the installation, data preparation and running. See GUIDELINES.md.

Using TAdaConv2d in your video backbone

To use TAdaConv2d in your video backbone, please follow the following steps:

# 1. copy models/module_zoo/ops/tadaconv.py somewhere in your project 
#    and import TAdaConv2d, RouteFuncMLP
from tadaconv import TAdaConv2d, RouteFuncMLP

class Model(nn.Module):
  def __init__(self):

    ...

    # 2. define tadaconv and the route func in your model
    self.conv_rf = RouteFuncMLP(
                c_in=64,            # number of input filters
                ratio=4,            # reduction ratio for MLP
                kernels=[3,3],      # list of temporal kernel sizes
    )
    self.conv = TAdaConv2d(
                in_channels     = 64,
                out_channels    = 64,
                kernel_size     = [1, 3, 3], # usually the temporal kernel size is fixed to be 1
                stride          = [1, 1, 1], # usually the temporal stride is fixed to be 1
                padding         = [0, 1, 1], # usually the temporal padding is fixed to be 0
                bias            = False,
                cal_dim         = "cin"
            )

     ...

  def self.forward(x):

    ...
    
    # 3. replace 'x = self.conv(x)' with the following line
    x = self.conv(x, self.conv_rf(x))

    ...

Initialization weight factorization. To use pre-trained weights of existing models, the weights for TAdaConv2d needs a bit factorization. The original shape of convolution weights Ci x Co x k x k needs to be expanded to 1 x 1 x Ci x Co x k x k. An option is to .unsqueeze(0) the weight twice. See convert_imagenet_weights in utils/checkpoint.py for more details.

Avoiding initializing layers that need to be skipped in initialization. The weights in the last layer in RouteFuncMLP is initialized as zeros, and those layers are marked with conv.skip=True. Make sure your codes do not alter the initial states of the RouteFuncMLP if you are to use it in your pre-trained models, by skipping the initialization for those convs as follows:

def your_initialization_function(model, ....):
  for m in model.modules():
    if hasattr(m, "skip_init") and m.skip_init:
      continue
    # your initialization codes next
    ...

Model Zoo

More of our pre-trained models are included in the MODEL_ZOO.md.

Feature Zoo

We include strong features for action localization on HACS and Epic-Kitchens-100 in our FEATURE_ZOO.md.

Contributors

This codebase is written and maintained by Ziyuan Huang, Zhiwu Qing and Xiang Wang.

Acknowledgement

Parts of the code are built upon SlowFast, timm, CoCLR, and BMN repositories.

Citations

If you find our codebase useful, please consider citing the respective work :).

@article{huang2023tadaconvv2,
  title={Temporally-Adaptive Models for Efficient Video Understanding},
  author={Huang, Ziyuan and Zhang, Shiwei and Pan, Liang and Qing, Zhiwu and Zhang, Yingya and Liu, Ziwei and Ang Jr, Marcelo H},
  journal={arXiv preprint arXiv:2308.05787},
  year={2023}
}

@inproceedings{huang2021tada,
  title={TAda! Temporally-Adaptive Convolutions for Video Understanding},
  author={Huang, Ziyuan and Zhang, Shiwei and Pan, Liang and Qing, Zhiwu and Tang, Mingqian and Liu, Ziwei and Ang Jr, Marcelo H},
  booktitle={{ICLR}},
  year={2022}
}

@inproceedings{mosi2021,
  title={Self-supervised motion learning from static images},
  author={Huang, Ziyuan and Zhang, Shiwei and Jiang, Jianwen and Tang, Mingqian and Jin, Rong and Ang, Marcelo H},
  booktitle={{CVPR}},
  pages={1276--1285},
  year={2021}
}

@article{huang2021towards,
  title={Towards training stronger video vision transformers for epic-kitchens-100 action recognition},
  author={Huang, Ziyuan and Qing, Zhiwu and Wang, Xiang and Feng, Yutong and Zhang, Shiwei and Jiang, Jianwen and Xia, Zhurong and Tang, Mingqian and Sang, Nong and Ang Jr, Marcelo H},
  journal={arXiv preprint arXiv:2106.05058},
  year={2021}
}

@article{qing2021stronger,
  title={A Stronger Baseline for Ego-Centric Action Detection},
  author={Qing, Zhiwu and Huang, Ziyuan and Wang, Xiang and Feng, Yutong and Zhang, Shiwei and Jiang, Jianwen and Tang, Mingqian and Gao, Changxin and Ang Jr, Marcelo H and Sang, Nong},
  journal={arXiv preprint arXiv:2106.06942},
  year={2021}
}