Research map / Learning theory and algorithms / Clustering
Video Face Clustering With Unknown Number of Clusters
Makarand Tapaswi, Marc T. Law, Sanja Fidler
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 |
|---|---|---|---|
| add L2 normalization of embeddings (ablation) | BCL · Validation (66 actors) | NMI: 74.57 -> 77.09 | Improved |
| train on single face images instead of averaging a random half subset of track images (ablation) | BCL · Validation (66 actors) | NMI: 7% lower NMI (absolute) with 124 clusters | Improved |
| use mini-batch size 2000 instead of 500 (ablation) | BCL · Validation (66 actors) | NMI: 72.13 -> 77.09 | Improved |
| use mini-batch size 2000 instead of 4000 (ablation) | BCL · Validation (66 actors) | WCP: 79.68 -> 85.65 | Improved |
| use embedding dimension 64 instead of 8 (ablation) | BCL · Validation (66 actors) | NMI: 47.62 -> 77.09 | Improved |
| use embedding dimension 64 instead of 32 (ablation) | BCL · Validation (66 actors) | NMI: 75.48 -> 77.09 | No clear effect |
| swap SE-ResNet50-256 for ResNet50 base CNN (ablation) | BCL · Validation (66 actors) | NMI: 77.09 -> 76.74 | No clear effect |
| add L2 normalization of embeddings (ablation) | Prototypical loss · Test episodes | NMI: over 15% NMI reduction | Hurt |
| replace prototypical loss with BCL loss, evaluated with HAC (K known) | clustering embeddings · Test BBT (6 episodes) | NMI: 68.3 -> 70.6 | Improved |
| replace prototypical loss with BCL loss, evaluated with K-means (K known) | clustering embeddings · Test BBT (6 episodes) | NMI: 60.7 -> 60.5 | Improved |
| fine-tune with automatically obtained positive/negative pairs | BCL · Test BBT (6 episodes) | NMI: 73.22 -> 88.26 | Improved |
| fine-tune with automatically obtained positive/negative pairs | BCL · Test BUFFY (6 episodes) | NMI: 71.23 -> 77.05 | Improved |
Closest papers by meaning
- VideoClusterNet: Self-Supervised and Adaptive Face Clustering for Videos (ECCV 2024 · similarity 0.91)
- Clustering Millions of Faces by Identity (TPAMI 2017 · similarity 0.90)
- End-To-End Face Detection and Cast Grouping in Movies Using Erdos-Renyi Clustering (ICCV 2017 · similarity 0.89)
- Self-supervised Video-centralised Transformer for Video Face Clustering (TPAMI 2023 · similarity 0.87)
- On Mitigating Hard Clusters for Face Clustering (ECCV 2022 · similarity 0.87)
- Linkage Based Face Clustering via Graph Convolution Network (CVPR 2019 · similarity 0.87)
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.
