Research map / Video / Video segmentation and tracking
Tracking Persons-of-Interest via Unsupervised Representation Adaptation
Shun Zhang, Jia‐Bin Huang, Jongwoo Lim, Yihong Gong, Jinjun Wang, Narendra Ahuja, Ming–Hsuan Yang
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 |
|---|---|---|---|
| use SymTriplet loss instead of triplet loss | Ours-Triplet · BBT05 clustering | weighted purity: 0.68 -> 0.85 | Improved |
| use SymTriplet loss instead of contrastive (Siamese) loss | Ours-Siamese · BBT05 clustering | weighted purity: 0.70 -> 0.85 | Improved |
| add contextual constraints to spatio-temporal constraints (ablation) | Ours-SymTriplet · BBT (multi-face tracking, overall) | MOTA (%): 57.2% -> 59.6% | Improved |
| add contextual constraints to spatio-temporal constraints (ablation) | Ours-SymTriplet · BBT05 clustering | weighted purity: 0.85 -> 0.86 | Improved |
| add linking of tracklets within each shot (ablation) | Ours-SymTriplet-Contx · BBT (multi-face tracking, overall) | MOTA (%): 43.1% -> 59.6% | Improved |
| add clustering of shot-level tracklets across shots (ablation) | Ours-SymTriplet-Contx · BBT (multi-face tracking, overall) | MOTA (%): 46.2% -> 59.6% | Improved |
| add clustering across shots (fewer ID switches) (ablation) | Ours-SymTriplet-Contx · BBT (multi-face tracking, overall) | ID switches (IDS): 1496 -> 817 | Improved |
| adapt the CNN features to the specific video instead of pre-trained features | pre-trained face features · BBT (multi-face tracking, overall) | MOTA (%): 30.0% -> 57.2% | Improved |
| nonlinear metric learning by fine-tuning all layers instead of Mahalanobis mapping on VGG-Face | VGG-Face-ULDML · T-ara clustering | weighted purity: 0.26 -> 0.84 | Improved |
| use two-threshold and Hungarian linking instead of mTLD for shot-level trajectories | Ours-mTLD · BBT (multi-face tracking, overall) | MOTA (%): 44.6% -> 57.2% | Improved |
| train features on one episode (BBT02) and transfer to others instead of per-video adaptation | Ours-SymTriplet · BBT05 clustering | weighted purity: 0.85 -> 0.79 | Hurt |
| use large feature dimensions (512, 1024) instead of 64 (ablation) | 64-d features · Bruno Mars clustering | weighted purity | Hurt |
Closest papers by meaning
- A Prior-Less Method for Multi-Face Tracking in Unconstrained Videos (CVPR 2018 · similarity 0.90)
- Face Association for Videos Using Conditional Random Fields and Max-Margin Markov Networks (TPAMI 2016 · similarity 0.88)
- Pre-training Strategies and Datasets for Facial Representation Learning (ECCV 2022 · similarity 0.86)
- Unsupervised Domain Adaptation for Face Recognition in Unlabeled Videos (ICCV 2017 · similarity 0.86)
- Joint Registration and Representation Learning for Unconstrained Face Identification (CVPR 2017 · similarity 0.85)
- VideoClusterNet: Self-Supervised and Adaptive Face Clustering for Videos (ECCV 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.
