Research map / Architectures and efficiency / CNN design
Virtual Fully-Connected Layer: Training a Large-Scale Face Recognition Dataset With Limited Computational Resources
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 classifier with Virtual FC | ResNet-50 face recognition · IJB-C@10-2 | TAR: 92.08 to 94.53 | Helped |
| Use Virtual FC instead of the metric-learning Multi-similarity method | ResNet-50 face recognition · IJB-C@10-2 | TAR: 90.06 to 94.53 | Helped |
| Increase samples per identity from K=2 to K=5 (ablation) | Virtual FC with CASIA-Net · LFW | accuracy: 98.60±0.48 -> 98.60±0.56 | No clear effect |
| Increase samples per identity from K=5 to K=10 (ablation) | Virtual FC with CASIA-Net · IJB-C | TAR: 92.50 -> 93.13 | Helped |
| Increase anchor count from M=0.01%×N to M=0.1%×N (ablation) | Virtual FC with CASIA-Net · LFW | accuracy: 94.67±1.26 -> 98.46±0.55 | Helped |
| Add free anchors by increasing M from 0.1%×N to 1%×N (ablation) | Virtual FC with CASIA-Net · IJB-C | TAR: 92.15 -> 92.21 | No clear effect |
| Increase anchor count from M=1%×N to M=10%×N (ablation) | Virtual FC with CASIA-Net · IJB-C | TAR: 92.21 -> 93.26 | Helped |
| Replace re-grouping with sampling that avoids same-group identities in a batch (ablation) | Virtual FC with ResNet-50 · IJB-C | TAR: 94.53 -> 92.05 | Hurt |
| Replace re-grouping with sampling that avoids same-group identities in a batch (ablation) | Virtual FC with ResNet-101 · IJB-C | TAR: 95.30 -> 89.60 | Hurt |
| Train with multi-task rather than single-task fine-tuning (ablation) | Virtual FC with ResNet-50 · LFW | accuracy: 98.23±0.98 -> 99.32±0.27 | Helped |
| Train Virtual FC as a single task from scratch rather than fine-tuning a pretrained model (ablation) | Virtual FC face recognition · IJB-C | TAR: 94.47 -> 89.37 | Hurt |
| Increase anchor count from M=10%×N to M=100%×N (ablation) | Virtual FC with CASIA-Net · MegaFace | rank1@106 accuracy: 82.29 -> 83.07 | Helped |
Closest papers by meaning
- Killing Two Birds With One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC (CVPR 2022 · similarity 0.90)
- An Efficient Training Approach for Very Large Scale Face Recognition (CVPR 2022 · similarity 0.89)
- Dynamic Class Queue for Large Scale Face Recognition in the Wild (CVPR 2021 · similarity 0.87)
- WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition (CVPR 2021 · similarity 0.87)
- Masked Face Recognition with Generative-to-Discriminative Representations (ICML 2024 · similarity 0.85)
- Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors (ICLR 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.