Code for training temporal fully-connected CRF models in Torch
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
pytorch speed improvement Aug 15, 2018
torch pytorch codebase release Aug 10, 2018 update readme Aug 10, 2018 First commit Jul 25, 2017

Asynchronous Temporal Fields for Activity Recognition Codebase

Contributor: Gunnar Atli Sigurdsson

  • This code implements a "Asynchronous Temporal Fields for Action Recognition" in Torch and PyTorch
  • This code extends the framework from

Details of the algorithm can be found in:

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)},
pdf = {},
code = {},

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 yields a final classification accuracy of 22.4% mAP (evaluated with charades_v1_classify.m).

Evaluation Scripts for localization and classification are available at

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

Baseline Codes for Activity Localization / Classification are available at