Research map / Learning theory and algorithms / Clustering
Clustering Millions of Faces by Identity
Charles Otto, Dayong Wang, Anil K. Jain
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 approximate rank-order clustering instead of original rank-order | face clustering · LFW (full) | F-measure: 0.80 -> 0.87 | Improved |
| use approximate rank-order instead of k-means | face clustering · LFW (full) | F-measure: 0.36 -> 0.87 | Improved |
| use randomized k-d tree approximate neighbors instead of brute-force (ablation) | approx. rank-order · LFW | F-measure: 0.72 -> 0.87 | Improved |
| use randomized k-d tree approximate neighbors instead of brute-force (ablation) | approx. rank-order · LFW + 1M | F-measure: 0.49 -> 0.79 | Improved |
| use randomized k-d tree instead of Chen et al. kNN graph (ablation) | approx. rank-order · LFW + 1M | F-measure: 0.41 -> 0.79 | Improved |
| use randomized k-d tree instead of brute-force neighbors (ablation) | approx. rank-order · LFW + 1M | run-time: 14:18:24 -> 00:07:20 | Improved |
| increase k-d tree search size linearly instead of fixed 2,000 (ablation) | approx. rank-order · LFW + 5M | F-measure: 0.13 -> 0.67 | Improved |
| increase search size logarithmically instead of linearly (ablation) | approx. rank-order · LFW + 5M | F-measure: 0.67 -> 0.33 | Hurt |
| add more distractor images (1M to 123M) | approx. rank-order · LFW + distractors | F-measure: 0.87 -> 0.27 | Hurt |
| exclude clusters of size 2 from cluster ranking (ablation) | per-cluster quality measure · LFW | correlation with precision: 0.427 -> 0.460 | Improved |
| combine intra and inter-cluster edge distance instead of inter alone (ablation) | per-cluster quality measure · LFW | correlation with precision: 0.325 -> 0.427 | Improved |
| use graph-based measures (coverage, MQ) instead of distance-based measures (ablation) | per-cluster quality measure · LFW | correlation with precision: 0.117 -> 0.325 | Hurt |
Closest papers by meaning
- On Mitigating Hard Clusters for Face Clustering (ECCV 2022 · similarity 0.91)
- End-To-End Face Detection and Cast Grouping in Movies Using Erdos-Renyi Clustering (ICCV 2017 · similarity 0.91)
- Structure-Aware Face Clustering on a Large-Scale Graph With 107 Nodes (CVPR 2021 · similarity 0.91)
- Learning to Cluster Faces on an Affinity Graph (CVPR 2019 · similarity 0.90)
- Video Face Clustering With Unknown Number of Clusters (ICCV 2019 · similarity 0.90)
- VideoClusterNet: Self-Supervised and Adaptive Face Clustering for Videos (ECCV 2024 · similarity 0.89)
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.
