Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Face Recognition from Hard Samples via Distribution Distillation Loss
Yuge Huang, Pengcheng Shen, Ying Tai, Shaoxin Li, Xiaoming Liu, Jilin Li, Feiyue Huang, Rongrong Ji
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 conventional finetuning with DDL | Arcface (CASIA+R50) · SCface average | rank-1 accuracy (%): 86.3 -> 94.4 | Improved |
| add order loss to KL loss (ablation) | DDL · SCface average | rank-1 accuracy (%): 89.6 -> 94.4 | Improved |
| add KL loss to order loss (ablation) | DDL · SCface average | rank-1 accuracy (%): 92.4 -> 94.4 | Improved |
| use KL divergence instead of EMD (ablation) | DDL · SCface average | rank-1 accuracy (%): 90.5 -> 94.4 | Improved |
| use KL divergence instead of JS divergence (ablation) | DDL · SCface average | rank-1 accuracy (%): 93.4 -> 94.4 | Improved |
| use hard negative mining instead of random negative selection (ablation) | DDL · SCface average | rank-1 accuracy (%): 90.6 -> 94.4 | Improved |
| use specific student distributions per resolution instead of one mixture (ablation) | DDL · SCface average | rank-1 accuracy (%): 92.1 -> 94.4 | Improved |
| mixture-trained DDL versus conventional finetuning | Arcface-FT · SCface average | rank-1 accuracy (%): 86.3 -> 92.1 | Improved |
| add DDL on the hardest low-resolution setting d1 | Arcface-FT (CASIA+R50) · SCface d1 | rank-1 accuracy (%): 67.3 -> 86.8 | Improved |
| add DDL on easy d3 images | Arcface-FT (CASIA+R50) · SCface d3 | rank-1 accuracy (%): 98.0 -> 98.3 | No clear effect |
| finetune with DDL on pose hard samples | Arcface (VGG+R100) · CPLFW | verification accuracy (%): 93.13 -> 93.43 | Improved |
| finetune with DDL on pose hard samples | Arcface (VGG+R100) · LFW | verification accuracy (%): 99.62 -> 99.68 | 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.
- Arcface — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · better on 3 of 3 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
- UniTSFace: Unified Threshold Integrated Sample-to-Sample Loss for Face Recognition (NeurIPS 2023 · similarity 0.90)
- AdaptiveFace: Adaptive Margin and Sampling for Face Recognition (CVPR 2019 · similarity 0.89)
- ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019 · similarity 0.89)
- ICD-Face: Intra-class Compactness Distillation for Face Recognition (ICCV 2023 · similarity 0.88)
- CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition (CVPR 2020 · similarity 0.88)
- EDFace-Celeb-1 M: Benchmarking Face Hallucination With a Million-Scale Dataset (TPAMI 2022 · 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.
