Research map / Label-efficient and robust learning / Recognition and long-tail
Dynamic Class Queue for Large Scale Face Recognition in the Wild
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the full learned classifier with a dynamic class queue using 10% of classes and generated weights | CosFace training on MF2 · MF2 identification | identification (%): 0.75% | Helped |
| Replace the full learned classifier with a dynamic class queue using 10% of classes and generated weights | CosFace training on MF2 · MF2 verification | verification (%): 1.72% | Helped |
| Remove tail classes and train only on classes with instances ≥ 9 (ablation) | CosFace on MF2 · MF2 identification | identification (%): 77.24 -> 77.10 | Hurt |
| Use class-based instead of instance-based sampling (ablation) | DCQ on MF2 · MF2 identification | identification (%): 77.99 -> 76.78 | Hurt |
| Use class-based instead of instance-based sampling (ablation) | DCQ on MF2 · MF2 verification | verification (%): 89.28 -> 88.44 | Hurt |
| Increase the queue from 2048 to 8192 classes (ablation) | DCQ on MS1MV2 · Face verification benchmarks | verification accuracy | Helped |
| Use 8192 queued classes, about 10% of all classes, rather than all classes (ablation) | DCQ on MS1MV2 · Face recognition benchmarks | recognition performance | No clear effect |
| Increase queue size beyond 8192 (ablation) | DCQ on MS1MV2 · CFP-FP and CPLFW | verification accuracy: within 0.3% | Hurt |
| Increase momentum α from 0.9 toward 0.999 (ablation) | DCQ on MS1MV2 · LFW, CFP-FP, AGEDB-30, and CPLFW | verification accuracy: over 6% | Helped |
| Replace the siamese one-shot weight generator with a prototypical network averaging support samples (ablation) | DCQ · Face recognition | performance | No clear effect |
| Use generated class weights with a 10% class queue rather than CosFace with all classes | Face recognition on MS1MV2 · LFW | accuracy (%): 99.78 -> 99.80 | No clear effect |
| Use generated class weights with a 10% class queue rather than CosFace with all classes | Face recognition on MS1MV2 · CALFW | accuracy (%): 96.20 -> 96.07 | Hurt |
Closest papers by meaning
- Accelerated Training for Massive Classification via Dynamic Class Selection (AAAI 2018 · similarity 0.91)
- An Efficient Training Approach for Very Large Scale Face Recognition (CVPR 2022 · similarity 0.90)
- HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere (ICLR 2025 · similarity 0.86)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.85)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.85)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 2025 · similarity 0.84)
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.