Long-Tail Temporal Action Segmentation with Group-wise Temporal Logit Adjustment

This repository is the official implementation of our paper "Long-Tail Temporal Action Segmentation via Group-wise Temporal Logit Adjustment" with model MSTCN on Breakfast dataset.

Dataset

The dataset used in our paper is open-source data. It can be downloaded from the references in the main paper.

Training

To train the models in the paper, run this command:

python activity_weighted_tla.py --action train --split 1 --seed 42 --w 0.5 --tau 0.5

You can find the scripts for other long-tailed methods in the code folder.

Evaluation

For evaluation, you need first generate the prediction for test set. For example,

python activity_weighted_tla.py --action predict --split 1 --seed 42 --w 0.5 --tau 0.5

Then, use the 'eval.py' script for evaluation. The results contain both global and balanced metrics.

python eval.py --method prediction_path