StillMix

This repo is the implementation of StillMix (ICCV2023): Mitigating and Evaluating Static Bias of Action Representations in the Background and the Foreground.

The codes are organized into three folders:

The benchmark_synthesis supports OOD benchmark synthesis.
The reference_network supports the training of the reference networks.
The main_network supports the training of the main networks.

Datasets

Training and IID evaluation

We use HMDB51, UCF101 and Kinetics400 in our training and IID evaluation. You can prepare these datasets following MMAction2.

We provide the dataset splits used in our paper on Google Drive.

OOD evaluation

We synthesize benchmarks for OOD evaluation. For details, please refer to our paper. We include the code for OOD benchmark synthesis in the benchmark_synthesis folder.

We release SCUBA and ConflFG on OneDrive. Since SCUFO is randomly sampled from SCUBA, we do not release SCUFO due to the limited storage space. Our experiments show that constructing SCOFO by sampling SCUBA with different random seeds only leads to slight performance difference (<1.0%) on SCUFO.

We provide the dataset splits used in our paper on Google Drive.

Reference Network

For training the reference network, please go to reference_network.

We save the results as pkl files. They are released on Google Drive.

Main Network

For training the main network, please go to main_network.

Citation

If you find our repository useful, please consider citing our paper.

@InProceedings{StillMix-2023,
author = {Haoxin Li and Yuan Liu and Hanwang Zhang and Boyang Li}, 
title = {Mitigating and Evaluating Static Bias of Action Representations in the Background and the Foreground}, 
booktitle = {International Conference on Computer Vision (ICCV)}, 
year = {2023}
}

Contact

If you have any problem please email me at haoxin003@e.ntu.edu.sg or lihaoxin05@gmail.com.

lihaoxin05/StillMix