Awesome
ACTION-Net
Official implementation of ACTION-Net: Multipath Excitation for Action Recognition (CVPR'21)
By Zhengwei Wang, Qi She and Aljosa Smolic
<p align="center"><img src="fig/backbone2.png" width="800" /></p> <p align="center"><img src="fig/heatmap_10_compressed.png" width="800" /></p>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)
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}
}