A pytorch-version implementation codes of paper: "Single Shot Temporal Action Detection", which is accepted in ACM MM 2017. paper
This repository is an improved version for the anchor-based part of A2Net.
If you find the repository helpful to you, here is a version that combines anchor-based module and anchor-free module: MSA-Net.
The detection results on THUMOS14 dataset:
mAP@ | 0.3 | 0.4 | 0.5 | 0.6 | 0.7 |
---|---|---|---|---|---|
61.83 | 58.03 | 51.12 | 35.93 | 20.72 |
This repository is implemented with Pytorch 1.1.0 + Python3.
The Two stream I3D feature could be downloaded from A2Net.
- To train the SSAD:
cd tools
python main.py
The parameters could be modified in
experiments\thumos\SSAD_train.yaml
- To Test the SSAD:
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