Research map / Video / Video segmentation and tracking
Self-supervised Video-centralised Transformer for Video Face Clustering
Yujiang Wang, Mingzhi Dong, Jie Shen, Yiming Luo, Yiming Lin, Pingchuan Ma, Stavros Petridis, Maja Pantić
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 |
|---|---|---|---|
| introduce video centres with naive options (uniform sampler, last-forward centres, incremental lr) (ablation) | CT TRSF. · EasyCom-Clustering V04 (known clusters) | NMI (%): 85.20 -> 83.54 | Hurt |
| use consecutive temporal sampler instead of uniform sampler (ablation) | VC TRSF. · EasyCom-Clustering V04 (known clusters) | NMI (%): 83.54 -> 83.97 | Improved |
| represent centres with incomplete centre algorithm instead of last-forward value (ablation) | VC TRSF. · EasyCom-Clustering V04 (known clusters) | NMI (%): 83.97 -> 86.20 | Improved |
| add periodic re-computation of video centres (complete algorithm) (ablation) | VC TRSF. · EasyCom-Clustering V04 (known clusters) | NMI (%): 86.20 -> 87.19 | Improved |
| make centre learning rate proportional to network learning rate (ablation) | VC TRSF. · EasyCom-Clustering V04 (known clusters) | NMI (%): 87.19 -> 90.69 | Improved |
| add scaled positional embedding (ablation) | VC TRSF. · EasyCom-Clustering V08 (known clusters) | NMI (%): 86.70 -> 84.53 | Hurt |
| add interpolated positional embedding (ablation) | VC TRSF. · EasyCom-Clustering V08 (known clusters) | NMI (%): 86.70 -> 85.89 | No clear effect |
| train transformer with video centres instead of pairwise contrast | CT TRSF. · BBT (known clusters) | NMI (%): 91.86 -> 94.20 | Improved |
| train transformer with video centres instead of pairwise contrast | CT TRSF. · BBT (unknown clusters) | cluster number difference: 76 -> 56 | Improved |
| replace TSiam with video-centralised transformer | TSiam · BBT (known clusters) | NMI (%): 94.11 -> 94.20 | No clear effect |
| learn a Siamese MLP (SSiam) instead of temporal averaging | Temp. AVG · EasyCom-Clustering (known clusters) | WCP (%): 78.44 -> 93.07 | Improved |
Closest papers by meaning
- VideoClusterNet: Self-Supervised and Adaptive Face Clustering for Videos (ECCV 2024 · similarity 0.89)
- Clustering Millions of Faces by Identity (TPAMI 2017 · similarity 0.87)
- Video Face Clustering With Unknown Number of Clusters (ICCV 2019 · similarity 0.87)
- On Mitigating Hard Clusters for Face Clustering (ECCV 2022 · similarity 0.86)
- End-To-End Face Detection and Cast Grouping in Movies Using Erdos-Renyi Clustering (ICCV 2017 · similarity 0.86)
- PrefAce: Face-Centric Pretraining with Self-Structure Aware Distillation (AAAI 2024 · similarity 0.84)
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.
