Research map / 3D vision / Face and shape
Towards High Fidelity Monocular Face Reconstruction With Rich Reflectance Using Self-Supervised Learning and Ray Tracing
Abdallah Dib, Cédric Thébault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, Louis Chevallier
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 9-band spherical harmonics instead of 3-band (ablation) | ray-traced face reconstruction · Face reconstruction (qualitative) | shadow estimation quality | Improved |
| use 9-band spherical harmonics instead of 7-band (ablation) | ray-traced face reconstruction · Face reconstruction (qualitative) | shadow estimation quality: slightly better | Improved |
| use two separate networks for diffuse and specular increments instead of one (ablation) | albedo increment estimation · Face reconstruction (qualitative) | albedo separation | Improved |
| add symmetry and consistency regularizers (ablation) | albedo personalization · Face reconstruction (qualitative) | residual shadows in albedo | Improved |
| train encoder jointly with the two albedo networks (ablation) | two-stage training · Face reconstruction (qualitative) | geometry fit | Improved |
| use differentiable ray tracing instead of a vertex-based renderer (ablation) | face reconstruction · Face reconstruction (qualitative) | shadow quality | Improved |
| increase ray tracing samples per pixel from 8 to 16 or 32 (ablation) | ray-traced training · training | reconstruction quality: 8, 16 and 32 spp | No clear effect |
| model self-geometry ray bounces (ablation) | ray-traced training · training | accuracy | No clear effect |
| use predicted attributes as initialization for optimization-based method | self-supervised network · Face reconstruction (qualitative) | reconstruction quality | Improved |
| ours versus best prior method on vertex position error | Ours · Geometry error on 23 GT-mesh images | vertex position error (cm): 0.174 vs 0.181 | Mixed |
| ours versus a noisy-mesh prior method on normal error | Ours · Geometry error on 23 GT-mesh images | normal error (rad): 0.159 -> 0.148 | Improved |
Closest papers by meaning
- Self-Supervised Multi-Level Face Model Learning for Monocular Reconstruction at Over 250 Hz (CVPR 2018 · similarity 0.90)
- Monocular Reconstruction of Neural Face Reflectance Fields (CVPR 2021 · similarity 0.90)
- Learning Complete 3D Morphable Face Models From Images and Videos (CVPR 2021 · similarity 0.89)
- Monocular Identity-Conditioned Facial Reflectance Reconstruction (CVPR 2024 · similarity 0.88)
- IntrinsicAvatar: Physically Based Inverse Rendering of Dynamic Humans from Monocular Videos via Explicit Ray Tracing (CVPR 2024 · similarity 0.87)
- Generative Multiview Relighting for 3D Reconstruction under Extreme Illumination Variation (CVPR 2025 · similarity 0.87)
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.
