Research map / 3D vision / Face and shape
Do 2D GANs Know 3D Shape? Unsupervised 3D Shape Reconstruction from 2D Image 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 |
|---|---|---|---|
| Use GAN2Shape rather than Unsup3d | BFM 3D shape reconstruction · BFM | SIDE (× 10^-2): 0.807 -> 0.756 | Helped |
| Use GAN2Shape rather than Unsup3d | BFM 3D shape reconstruction · BFM | MAD (deg.): 16.34 -> 14.81 | Helped |
| Add latent-offset regularization (ablation) | GAN2Shape · BFM | SIDE (× 10^-2): 0.925 -> 0.756 | Helped |
| Add latent-offset regularization (ablation) | GAN2Shape · BFM | MAD (deg.): 16.42 -> 14.81 | Helped |
| Train without the symmetry assumption instead of using Unsup3d | BFM 3D shape reconstruction · BFM without symmetry assumption | SIDE (× 10^-2): 1.334 -> 1.023 | Helped |
| Train without the symmetry assumption instead of using Unsup3d | BFM 3D shape reconstruction · BFM without symmetry assumption | MAD (deg.): 33.79 -> 17.09 | Helped |
| Render face rotation from recovered 3D shapes rather than the strongest listed baseline | Identity-preserving face rotation · Face rotation | error_mean (deg.): 41.17 -> 28.93 | Helped |
| Render face rotation from recovered 3D shapes rather than the strongest listed baseline | Identity-preserving face rotation · Face rotation | error_max (deg.): 58.93 -> 43.02 | Helped |
| Use a flat rather than ellipsoid shape prior (ablation) | GAN2Shape · BFM | SIDE (× 10^-2): 0.756 -> 1.021 | Hurt |
| Use a flat rather than ellipsoid shape prior (ablation) | GAN2Shape · BFM | MAD (deg.): 14.81 -> 20.46 | Hurt |
| Vary the ellipsoid prior to asymmetric, shifted, or weaker variants (ablation) | GAN2Shape · BFM | SIDE (× 10^-2): 0.756; 0.769; 0.767; 0.775; 0.764 | No clear effect |
| Reduce GAN training data from 160k to 3k samples (ablation) | GAN2Shape · BFM | SIDE (× 10^-2): 0.756 -> 0.794 | No clear effect |
Closest papers by meaning
- Unsupervised 3D Shape Completion Through GAN Inversion (CVPR 2021 · similarity 0.91)
- 3D GAN Inversion With Facial Symmetry Prior (CVPR 2023 · similarity 0.89)
- 3D generation on ImageNet (ICLR 2023 · similarity 0.89)
- A Geometric Analysis of Deep Generative Image Models and Its Applications (ICLR 2021 · similarity 0.89)
- DECOR-GAN: 3D Shape Detailization by Conditional Refinement (CVPR 2021 · similarity 0.89)
- What You See is What You GAN: Rendering Every Pixel for High-Fidelity Geometry in 3D GANs (CVPR 2024 · similarity 0.89)
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.