Research map / Label-efficient and robust learning / Semi-supervised learning
3FabRec: Fast Few-Shot Face Alignment by Reconstruction
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 |
|---|---|---|---|
| Fine-tune the encoder on landmark errors (ablation) | 3FabRec · 300-W full testset | NME (%): 4.16 -> 3.82 | Helped |
| Fine-tune the encoder on landmark errors (ablation) | 3FabRec · AFLW full testset | NME (%): 2.12 -> 1.84 | Helped |
| Fine-tune the encoder on landmark errors (ablation) | 3FabRec · WFLW full testset | NME (%): 6.11 -> 5.62 | Helped |
| Add the two lower ITL layers to layers 3+4 (ablation) | 3FabRec · 300-W, outlines only | NME: 18.17 -> 6.58 | Helped |
| Add adversarial loss to the autoencoder (ablation) | 3FabRec autoencoder · 300-W full testset | NME (%): 5.67 -> 5.30 | Helped |
| Add structural image loss to the adversarial autoencoder (ablation) | 3FabRec autoencoder · 300-W full testset | NME (%): 5.30 -> 4.92 | Helped |
| Fine-tune the encoder after adding adversarial and structural losses (ablation) | 3FabRec autoencoder · 300-W full testset | NME (%): 4.92 -> 4.41 | Helped |
| Pretrain the autoencoder on 2.1M unlabeled face images instead of 100k (ablation) | C-FT · 300-W full testset | NME: 4.73 -> 4.41 | Helped |
| Use 2.1M-image pretraining rather than no external pretraining (ablation) | C-FT · 300-W with 1.5% labeled data | NME: 7.95 -> 5.71 | Helped |
| Reduce the embedding size from 99 to 50 dimensions (ablation) | 3FabRec · 300-W full testset | NME (%): 4.41 -> 4.59 | Hurt |
| Reduce the embedding size from 99 to 50 dimensions (ablation) | 3FabRec · 300-W full testset | FR@0.1 (%): 1.45 -> 1.02 | Helped |
| Use the pretrained low-dimensional face representation rather than a ResNet-18 without pretraining (ablation) | 3FabRec C-FT · 300-W with 1.5% labeled data | NME: 8.70 -> 5.71 | Helped |
Closest papers by meaning
- RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild (CVPR 2020 · similarity 0.88)
- Learning to Detect 3D Facial Landmarks via Heatmap Regression with Graph Convolutional Network (AAAI 2022 · similarity 0.88)
- Brute-Force Facial Landmark Analysis With a 140,000-Way Classifier (AAAI 2018 · similarity 0.87)
- FaceLift: Semi-supervised 3D Facial Landmark Localization (CVPR 2024 · similarity 0.86)
- T-FAKE: Synthesizing Thermal Images for Facial Landmarking (CVPR 2025 · similarity 0.85)
- LAFS: Landmark-based Facial Self-supervised Learning for Face Recognition (CVPR 2024 · 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.