Research map / Image generation / Super-resolution
Super-FAN: Integrated Facial Landmark Localization and Super-Resolution of Real-World Low Resolution Faces in Arbitrary Poses With GANs
Read the paperCheck 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 |
|---|---|---|---|
| jointly train the alignment and super-resolution networks end-to-end (ablation) | Super-FAN · LS3D-W balanced | AUC: 61.0% -> 67.0% (pose [0-30]); 55.6% -> 63.0% (pose [30-60]); 42.3% -> 52.5% (pose [60-90]) | Helped |
| add the heatmap loss (ablation) | Super-resolution network with pixel and perceptual losses · LS3D-W balanced | AUC: 50.2% -> 55.6% (pose [30-60]) | Helped |
| add the perceptual loss (ablation) | Super-resolution network with pixel loss · LS3D-W balanced | AUC: 45.3% -> 50.2% (pose [30-60]) | Helped |
| add the GAN loss and end-to-end training (ablation) | Ours-pixel-feature-heatmap · LS3D-W balanced | AUC: 61.0% -> 67.0% (pose [0-30]); 55.6% -> 63.0% (pose [30-60]); 42.3% -> 52.5% (pose [60-90]) | Helped |
| use the proposed residual-based super-resolution architecture instead of SR-ResNet | Super-resolution followed by FAN · LS3D-W balanced | AUC: 48.9% -> 57.0% (pose [0-30]); 38.9% -> 50.2% (pose [30-60]); 21.4% -> 34.9% (pose [60-90]) | Helped |
| add more blocks at higher resolutions using the 12-3-2 distribution instead of 16-1-1 (ablation) | Super-resolution network · LS3D-W balanced | visual detail quality | Helped |
| remove the long skip connection (ablation) | Super-resolution network · Super-resolution experiments | performance | No clear effect |
| replace PReLU with ReLU (ablation) | Super-resolution network · Super-resolution experiments | performance | No clear effect |
| use WGAN loss instead of the GAN loss formulation used in prior work (ablation) | Super-resolution training · Training stability | training stability | Helped |
| add the GAN loss to the pixel-feature-heatmap model (ablation) | Super-resolution quality measured by PSNR and SSIM · LS3D-W balanced | PSNR: 21.55 -> 20.85 (pose [0-30]); 22.55 -> 21.67 (pose [30-60]); 23.17 -> 22.24 (pose [60-90]) | Hurt |
| retrain Super-FAN with random Gaussian blur, simulated JPEG artifacts, and color distortion (ablation) | Super-FAN · WiderFace real-world low-resolution faces | visual quality | Helped |
Closest papers by meaning
- FSRNet: End-to-End Learning Face Super-Resolution With Facial Priors (CVPR 2018 · similarity 0.91)
- Joint Super-Resolution and Alignment of Tiny Faces (AAAI 2020 · similarity 0.91)
- Deep Face Super-Resolution With Iterative Collaboration Between Attentive Recovery and Landmark Estimation (CVPR 2020 · similarity 0.89)
- FAN-Face: a Simple Orthogonal Improvement to Deep Face Recognition (AAAI 2020 · similarity 0.87)
- POPoS: Improving Efficient and Robust Facial Landmark Detection with Parallel Optimal Position Search (AAAI 2025 · similarity 0.85)
- Low-Light Face Super-resolution via Illumination, Structure, and Texture Associated Representation (AAAI 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.