ACTION-Net

Official implementation of ACTION-Net: Multipath Excitation for Action Recognition (CVPR'21)

By Zhengwei Wang, Qi She and Aljosa Smolic

Getting Started

• EgoGesture data folder structure

  |-frames
  |---Subject01
  |------Scene1
  |---------Color1
  |------------rgb1
  |---------------000001.jpg
  ......
  |-labels
  |---Subject01
  |------Scene1
  |---------Group1.csv
  ......

• Something-Something V2

  |-frames
  |---1
  |------000001.jpg
  |------000002.jpg
  |------000003.jpg
  ......

• Jester

  |-frames
  |---1
  |------000001.jpg
  |------000002.jpg
  |------000003.jpg
  ......

  Requirements

  Provided in the action.Dockerfile

Annotation files

Annotation files are at this link. Please follow the annotation files to construct the frame path.

Usage

sh train_ego_8f.sh 0,1,2,3 if you use four gpus

Acknowledgment

Our codes are built based on previous repos TSN, TSM and TEA

Pretrained models

Currently, we do not provide the pretrained models since we reconstruct the structure and rename our modules of ACTION for public release. It should be able to get the similar performance indicated in the paper using the codes provided above.

(Update)

EgoGesture using 8f

Jester using 8f

Citation

If you find our work useful in your research, please cite:

@InProceedings{Wang_2021_CVPR,
author = {Wang, Zhengwei and She, Qi and Smolic, Aljosa},
title = {ACTION-Net: Multipath Excitation for Action Recognition},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021}
}