Research map / Label-efficient and robust learning / Recognition and long-tail
Towards Robust Face Recognition with Comprehensive Search
Manyuan Zhang, Guanglu Song, Yu Liu, Hongsheng Li
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 data-cleaning search to loss and backbone search (ablation) | loss+backbone search · IJB-B TAR@FAR=1e-5, MS1MV2 | TAR (%): 83.67 -> 85.66 | Improved |
| add data-cleaning search to loss and backbone search (ablation) | loss+backbone search · IJB-C TAR@FAR=1e-5, MS1MV2 | TAR (%): 89.55 -> 90.45 | Improved |
| search data cleaning only instead of hand-crafted cleaning (ablation) | baseline · IJB-C TAR@FAR=1e-5, MS1MV2 | TAR (%): 87.29 -> 88.47 | Improved |
| search loss scales and margins only (ablation) | ArcFace baseline · IJB-C TAR@FAR=1e-5, MS1MV2 | TAR (%): 87.29 -> 88.88 | Improved |
| search backbone width and depth only (ablation) | modified MobileNet · IJB-C TAR@FAR=1e-5, MS1MV2 | TAR (%): 87.29 -> 87.92 | Improved |
| search all three jointly instead of each alone (ablation) | best single-aspect search · IJB-B TAR@FAR=1e-5, MS1MV2 | TAR (%): 83.26 -> 85.66 | Improved |
| joint search vs hand-crafted baseline | baseline · IJB-C TAR@FAR=1e-4, MS1MV2 | TAR (%): 92.61 -> 93.58 | Improved |
| joint search vs hand-crafted baseline | baseline · IJB-C TAR@FAR=1e-5, Glint360K | TAR (%): 86.33 -> 90.06 | Improved |
| joint search vs hand-crafted baseline | baseline · IJB-B TAR@FAR=1e-5, Glint360K | TAR (%): 79.83 -> 86.06 | Improved |
| use features from Glint360K model instead of MS1MV2 model for cleaning (ablation) | data search · IJB-C TAR@FAR=1e-5 | TAR (%): 88.80 -> 88.84 | No clear effect |
| train with combination searched on Glint360K instead of on MS1MV2 (ablation) | joint search · IJB-C TAR@FAR=1e-5, MS1MV2 train | TAR (%): 90.45 -> 89.41 | Hurt |
| transferred Glint360K-searched combination vs hand-crafted baseline | baseline · IJB-C TAR@FAR=1e-5, MS1MV2 train | TAR (%): 87.29 -> 89.41 | Improved |
Closest papers by meaning
- Loss Function Search for Face Recognition (ICML 2020 · similarity 0.90)
- BroadFace: Looking at Tens of Thousands of People at Once for Face Recognition (ECCV 2020 · similarity 0.87)
- AdaptiveFace: Adaptive Margin and Sampling for Face Recognition (CVPR 2019 · similarity 0.87)
- Loss Function Discovery for Object Detection via Convergence-Simulation Driven Search (ICLR 2021 · similarity 0.86)
- UniTSFace: Unified Threshold Integrated Sample-to-Sample Loss for Face Recognition (NeurIPS 2023 · similarity 0.86)
- DamoFD: Digging into Backbone Design on Face Detection (ICLR 2023 · similarity 0.86)
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.
