A temporal action detection approach named MSA-Net that combines anchor-based and anchor-free modules.
This repository is an improved version of A2Net, the code for anchor-free module references FCOS.
If you find the repository helpful to you, here is the anchor-based part SSAD.
The detection results on THUMOS14 dataset:
| mAP@ | 0.3 | 0.4 | 0.5 | 0.6 | 0.7 |
|---|---|---|---|---|---|
| 62.79 | 58.87 | 51.47 | 38.75 | 24.03 |
This repository is implemented with Pytorch 1.1.0 + Python3.
The Two stream I3D feature could be downloaded from A2Net.
- To train the MSA-Net:
cd tools
python main.py
The parameters could be modified in
experiments\A2Net_thumos.yaml
- To Test the MSA-Net:
cd tools
python eval.py --checkpoint $cpt_path
- Evaluating the detection performance:
Open Matlab in lib\Evaluation\THUMOS14_evalkit_20150930 path, and put the testing result file in the path, and execute the file:
multi_iou_eval