Research map / Graph learning / Graph algorithms and clustering
Enhancing Noise Resilience in Face Clustering via Sparse Differential Transformer
Dafeng Zhang, Yongqi Song, Shizhuo Liu
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 vanilla Transformer with Differential Transformer (ablation) | Face clustering predictor · MS1M 584K unlabeled | Pairwise F-score: 94.25 -> 95.05 | Improved |
| add Top-K attention mask to Diff Transformer (ablation) | Diff Transformer · MS1M 584K unlabeled | Pairwise F-score: 95.05 -> 95.34 | Improved |
| add Top-K attention mask to vanilla Transformer (ablation) | Vanilla Transformer · MS1M 584K unlabeled | Pairwise F-score: 94.25 -> 94.78 | Improved |
| add MoE over Top-(K-u), Top-K, Top-(K+u) masks (ablation) | SDT · MS1M 584K unlabeled | Pairwise F-score: 95.34 -> 95.46 | Improved |
| replace exponential distance transform with sigmoid (delta=7.5) (ablation) | FC-ESER · MS1M 584K unlabeled | Pairwise F-score: 94.49 -> 94.67 | Improved |
| use sigmoid steepness delta=5.0 instead of exp transform (ablation) | FC-ESER · MS1M 584K unlabeled | Pairwise F-score: 94.49 -> 94.09 | Hurt |
| use sigmoid steepness delta=10.0 instead of 7.5 (ablation) | FC-ESER · MS1M 584K unlabeled | Pairwise F-score: 94.67 -> 94.45 | Hurt |
| add Top-K Jaccard restricted to predicted neighbors (ablation) | FC-ESER + Eq. (6) · MS1M 584K unlabeled | Pairwise F-score: 94.67 -> 95.02 | Improved |
| replace vanilla with Diff Transformer in LCEPCE (ablation) | LCEPCE · DeepFashion | Pairwise F-score: 38.43 -> 41.76 | Improved |
| replace vanilla with Diff Transformer in LCEPCE (ablation) | LCEPCE · MS1M 584K | Pairwise F-score: 93.43 -> 94.68 | Improved |
| vary score threshold eta from 0.5 to 0.9 (ablation) | MoE-SDT · MS1M 584K unlabeled | Pairwise F-score: 95.34 -> 95.46 | No clear effect |
| vary kNN graph size K (40, 80, 160) (ablation) | MoE-SDT · MS1M 584K unlabeled | Pairwise F-score: 95.32 / 95.46 / 95.34 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Ada-NETS — Ada-NETS: Face Clustering via Adaptive Neighbour Discovery in the Structure Space (ICLR 2022) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- On Mitigating Hard Clusters for Face Clustering (ECCV 2022 · similarity 0.89)
- Structure-Aware Face Clustering on a Large-Scale Graph With 107 Nodes (CVPR 2021 · similarity 0.87)
- STAR-FC: Structure-Aware Face Clustering on Ultra-Large-Scale Graphs (TPAMI 2023 · similarity 0.86)
- Linkage Based Face Clustering via Graph Convolution Network (CVPR 2019 · similarity 0.86)
- Local Connectivity-Based Density Estimation for Face Clustering (CVPR 2023 · similarity 0.86)
- TopoFR: A Closer Look at Topology Alignment on Face Recognition (NeurIPS 2024 · similarity 0.83)
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.
