Contributor: Gunnar Atli Sigurdsson

• This code implements a "Asynchronous Temporal Fields for Action Recognition" in Torch and PyTorch
• This code extends the framework from github.com/gsig/charades-algorithms

Details of the algorithm can be found in:

@inproceedings{sigurdsson2017asynchronous,
author = {Gunnar A. Sigurdsson and Santosh Divvala and Ali Farhadi and Abhinav Gupta},
title = {Asynchronous Temporal Fields for Action Recognition},
booktitle={The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2017},
pdf = {https://arxiv.org/pdf/1612.06371.pdf},
code = {https://github.com/gsig/temporal-fields},
}

We have updated the codebase with an improved and simplified PyTorch model. Detail can be found under pytorch

Using the improved PyTorch code, a simple RGB model obtains 26.1% mAP (evaluated with charades_v1_classify.m). Using the original Torch code, combining the predictions (submission files) of those models using combine_rgb_flow.py yields a final classification accuracy of 22.4% mAP (evaluated with charades_v1_classify.m).

Evaluation Scripts for localization and classification are available at allenai.org/plato/charades/

Submission files for temporal fields and baselines for classification and localization that are compatible with the official evaluation codes on allenai.org/plato/charades/ are available here: charades_submission_files.zip. This might be helpful for comparing and contrasting different algorithms.

Baseline Codes for Activity Localization / Classification are available at github.com/gsig/charades-algorithms