This repository contains the codebase for Anticipation Transformer (Anticipatr) proposed in the ECCV'22 paper.
Our method proposes a two-stage training method for the task of long-term action anticipation.
-
pretrainingdirectory contains code for stage 1. This stage involves training a model for the task of snippet-based action anticipation. -
srcdirectory contains code for stage 2. This stage uses the frozen networks from stage 1 and trains two other networks for the task of long-term action anticipation. -
The file
env.ymlprovides dependencies for the training environment -
To train the model, download the features to the path given in the .py file or change the path to a custom location.
For our training setup, we use these directories.
-
pretraining_dataprovides the data specific to our training for stage 1. -
For our implementation, we save data in a directory
datawithin this top-level directory. For training the data needs to be downloaded from cited sources.- The features and annotations for Epic-Kitchens-55 and EGTEA can be downloaded from https://github.com/facebookresearch/ego-topo
- The features and annotations for Breakfast can be downloaded from https://github.com/yabufarha/ms-tcn
-
We also add sample run scripts for the codes in the directory
run_scripts
@inproceedings{nawhal2022anticipatr,
title={Rethinking Learning Approaches for Long-Term Action Anticipation},
author={Nawhal, Megha and Jyothi, Akash Abdu and Mori, Greg},
booktitle={Proceedings of the European Conference on Computer Vision},
year={2022}
}
For further questions, please email Megha Nawhal.