Context-Enhanced Memory-Refined Transformer for Online Action Detection

May 17, 2025 ยท View on GitHub

This is an implementation for our CVPR 2025 paper "Context-Enhanced Memory-Refined Transformer for Online Action Detection".

Data Preparation

You can directly download the pre-extracted feature of THUMOS and EK100 from TeSTra.

You can find the pre-extracted feature of CrossTask here

Training

You can train our method on THUMOS'14 by simply running:

python main.py --config_file configs/THUMOS/cmert_long256_work4_kinetics_1x.yaml

Evaluation

You can verify the performance of our method on THUMOS'14 by simply running:

python main.py --test 1 --config_file configs/THUMOS/cmert_long256_work4_kinetics_1x.yaml MODEL.CHECKPOINT checkpoints/THUMOS/cmert_long256_work4_kinetics_1x/epoch-9.pth MODEL.LSTR.INFERENCE_MODE batch

This should return the results reported in the submitted paper.

Acknowledgements

This codebase is built upon TeSTra and MAT.