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