Research map / Label-efficient and robust learning / Recognition and long-tail
Caption-Supervised Face Recognition: Training a State-of-the-Art Face Model without Manual Annotation
Qingqiu Huang, Lei Yang, Huaiyi Huang, Tong Wu, Dahua Lin
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 propagation data (S2) to One2One data (ablation) | CSFR · MegaFace | top-1 identification (%): 79.81 -> 82.41 | Improved |
| add expansion data (S3) to One2One plus propagation data (ablation) | CSFR · MegaFace | top-1 identification (%): 82.41 -> 83.34 | Improved |
| add propagation data (S2) to One2One (ablation) | CSFR · IJB-A | TAR@FAR=0.001 (%): 68.67 -> 75.06 | Improved |
| add expansion data (S3) to One2One plus propagation (ablation) | CSFR · IJB-A | TAR@FAR=0.001 (%): 75.06 -> 77.50 | Improved |
| train on automatically annotated data only (S2+S3), dropping One2One (ablation) | CSFR · MegaFace | top-1 identification (%): 69.52 vs 79.81 | No clear effect |
| add K-means clustering of unlabeled faces to One2One | S1 baseline · MegaFace | top-1 identification (%): 79.81 -> 79.9 | No clear effect |
| add LTC clustering of unlabeled faces to One2One | S1 baseline · IJB-A | TAR@FAR=0.001 (%): 68.67 -> 71.34 | Improved |
| use max-pooling instead of mean-pooling in MIL (ablation) | MIL · MegaFace | top-1 identification (%): 69.62 -> 73.11 | Improved |
| use CSFR instead of best MIL method (max-pooling) | CSFR · IJB-A | TAR@FAR=0.001 (%): 49.06 -> 77.50 | Improved |
| train ArcFace on MovieFace instead of MS1M | ArcFace · IJB-A | TAR@FAR=0.001 (%): 96.82 -> 96.96 | Improved |
| train ArcFace on MovieFace instead of MS1M | ArcFace · MegaFace | top-1 identification (%): 97.92 -> 96.96 | Hurt |
| train ArcFace on MovieFace instead of IMDb-Face | ArcFace · IJB-A | TAR@FAR=0.001 (%): 94.25 -> 96.96 | Improved |
Closest papers by meaning
- Partially-Supervised Image Captioning (NeurIPS 2018 · similarity 0.88)
- Learning to Overcome Noise in Weak Caption Supervision for Object Detection (TPAMI 2022 · similarity 0.86)
- Beyond a Pre-Trained Object Detector: Cross-Modal Textual and Visual Context for Image Captioning (CVPR 2022 · similarity 0.86)
- nocaps: novel object captioning at scale (ICCV 2019 · similarity 0.86)
- Pre-training Strategies and Datasets for Facial Representation Learning (ECCV 2022 · similarity 0.86)
- GroundingFace: Fine-grained Face Understanding via Pixel Grounding Multimodal Large Language Model (CVPR 2025 · similarity 0.85)
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.
