Research map / 3D vision / Human pose estimation
VPN++: Rethinking Video-Pose Embeddings for Understanding Activities of Daily Living
Srijan Das, Rui Dai, Di Yang, François Brémond
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace MSE with supervised contrastive distillation (SCD) loss (ablation) | VPN-F · Smarthome CS | top-1 accuracy: 61.8 -> 67.1 | Improved |
| replace CRD with SCD loss (ablation) | VPN-F · Smarthome CS | top-1 accuracy: 64.7 -> 67.1 | Improved |
| replace CRD with SCD loss (ablation) | VPN-F · NTU-60 CS | top-1 accuracy: 90.1 -> 90.8 | Improved |
| use pose-driven attention instead of RGB Non-Local attention (ablation) | I3D · Smarthome CS | top-1 accuracy: 53.6 -> 65.2 | Improved |
| add RGB Non-Local attention to I3D (ablation) | I3D · Smarthome CS | top-1 accuracy: 53.4 -> 53.6 | No clear effect |
| use MSE on attention weights instead of SCD at feature level (ablation) | VPN-A · Smarthome CS | top-1 accuracy: 55.7 -> 62.1 | Improved |
| add collaborative training of teacher and student for MSE attention distillation (ablation) | VPN-A · Smarthome CS | top-1 accuracy: 61.1 -> 62.1 | Improved |
| add collaborative training for SCD attention-weight distillation (ablation) | VPN-A · Smarthome CS | top-1 accuracy: 54.2 -> 53.1 | Hurt |
| combine VPN-F and VPN-A in one multi-teacher model instead of sequentially (ablation) | VPN++ · Smarthome CS | top-1 accuracy: 62.3 -> 69.0 | Improved |
| use multi-teacher VPN++ instead of late fusion of VPN-F and VPN-A (ablation) | VPN++ · Smarthome CS | top-1 accuracy: 68.7 -> 69.0 | Improved |
| add spatial embedding L_e to the attention teacher (ablation) | VPN++ · Smarthome CS | top-1 accuracy: 68.9 -> 69.0 without L_e | No clear effect |
| distill poses into RGB (VPN++) instead of using poses at test time (VPN), poor pose quality (ablation) | VPN++ · Smarthome CS (low-quality poses) | top-1 accuracy: 62.1 -> 66.8 | Improved |
Closest papers by meaning
- VPN: Learning Video-Pose Embedding for Activities of Daily Living (ECCV 2020 · similarity 0.94)
- Just Add ?! Pose Induced Video Transformers for Understanding Activities of Daily Living (CVPR 2024 · similarity 0.88)
- ViPNAS: Efficient Video Pose Estimation via Neural Architecture Search (CVPR 2021 · similarity 0.86)
- Toyota Smarthome: Real-World Activities of Daily Living (ICCV 2019 · similarity 0.86)
- RPAN: An End-To-End Recurrent Pose-Attention Network for Action Recognition in Videos (ICCV 2017 · similarity 0.85)
- FreeMan: Towards Benchmarking 3D Human Pose Estimation under Real-World Conditions (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
