Awesome

Structured Attention Composition for Temporal Action Localization

This repository is the official implementation of SAC. In this work, we tackle the temporal action localization task from the perspective of modality, and precisely assign frame-modality attention. Paper from arXiv or IEEE.

Requirements

To install requirements:

conda env create -n env_name -f environment.yaml

Before running the code, please activate this conda environment.

Data Preparation

a. Download pre-extracted features from baiduyun (code:6666)

Please ensure the data structure is as below

├── data
   └── thumos
       └── val
           ├── video_validation_0000051_02432.npz
           ├── video_validation_0000051_02560.npz
           ├── ...
       └── test
           ├── video_test_0000004_00000.npz
           ├── video_test_0000004_00256.npz
           ├── ...

Train

a. Config

Adjust configurations.

./experiments/thumos/network.yaml

c. Train

cd tools
bash run.sh

Inference

a. You can download pre-trained models from baiduyun (code:6666), and put the weight file in the folder checkpoint.

• Performance

b. Test

cd tools
python eval.py

Related Projects

• BackTAL: Background-Click Supervision for Temporal Action Localization.
• A2Net: Revisiting Anchor Mechanisms for Temporal Action Localization.

Contact

For any discussions, please contact nwpuyangle@gmail.com.