Improvements
- AdaTAD's results on Multi-THUMOS and Charades are released at here. We achieve an amazing performance of 48.73% average mAP on Multi-THUMOS dataset and 37.56% average mAP on Charades dataset.
Bug Fixes
- The annotation of MultiTHUMOS dataset is updated. We fix few wrong annotations, and manually remove those actions longer than the video duration. The corresponding results of ActionFormer and TemporalMaxer are also updated, which are very close to the previous results.
- The annotation of Charades dataset is updated. We remove the empty annotations and add two missing videos. This annotation is aligned with the TriDet's implementation now. The corresponding result of ActionFormer is updated.
New features
- OpenTAD supports temporal checkpointing now, which is proposed in ETAD and can save more memory during end-to-end training. Check here.
- We support Ego4D-MQ dataset now. Please check here to download the annotation and features, and here for an implementation on ActionFormer.
- We release the config and pretrained checkpoints of AdaTAD on EPIC-KITCHENS and Ego4D datasets. Please check here.
Improvements
- VSGN now supports THUMOS dataset.
- ETAD now supports end-to-end training.
- The side-tuning of AdaTAD is supported now. Please check here.
- AdaTAD is released with configs / logs / pretrained checkpoints on ActivityNet dataset and THUMOS14 dataset. Please check AdaTAD.
- We release non-reversible version of Re2TAL, as a baseline of end-to-end actionformer. Please check Re2TAL.
- Fix typos.
- Initial release of OpenTAD.
- The codebase is under beta testing. Feedbacks and suggestions are welcome!