official code for PseR: Pseudo-label Refinement for Point-Supervised Temporal Action Detection
- Install the python environment and get dataset of HR-Pro
cd models/opsand run:python setup.py build_ext --inplacecd ../..
PseR generates seed proposals based on existing methods. Take LACP, for example
- Run LACP to get the initial prediction
- Execute
python seed_process/ge_seed_proposal.pyto getlacp_seed_final.json - [Optionally], you can use the lacp_seed_final.json[7vx4] we have already gotten
- Run
python main.pyto get the pseudo-label of the PseR prediction:lacp_pser.json - [Optionally], you can use the lacp_pser.json[ip52] we have already gotten
We trained on the THUMOS'14 dataset based on OpenTAD and lacp_pser.json
ActF stands for ActionFormer
| Model | mAP@0.3 | mAP@0.4 | mAP@0.5 | mAP@0.6 | mAP@0.7 | ave. mAP | Config | Download |
|---|---|---|---|---|---|---|---|---|
| LACP | 64.6 | 56.5 | 45.3 | 34.5 | 21.8 | 44.5 | ||
| LACP+ActF | 77.1 | 68.5 | 56.2 | 41.8 | 23.3 | 53.4 | config | log[k7b4] | model |
| Ours+ActF | 78.5 | 70.9 | 59.3 | 43.2 | 26.2 | 55.6 | config | log[a67a] | model |
Our code is based on HR-Pro, OpenTAD, LACP. We would like to express our gratitude for their outstanding work