Research map / 3D vision / Depth and stereo
Registration-Free Learnable Multi-View Capture of Faces in Dense Semantic Correspondence
Panagiotis P. Filntisis, George Retsinas, Radek Danecek, Vanessa Sklyarova, Petros Maragos, Timo Bolkart
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 |
|---|---|---|---|
| train without precomputed registrations (MOCHI vs TEMPEH trained with registrations) | TEMPEH · FaMoS test | median point-to-surface error (mm): 0.36 -> 0.26 | Improved |
| zero-shot transfer to CoMA (MOCHI vs TEMPEH) | TEMPEH · CoMA zero-shot | median point-to-surface error (mm): 0.81 -> 0.53 | Improved |
| add test-time optimization (50 iterations) with scan | MOCHI · FaMoS test | median point-to-surface error (mm): 0.26 -> 0.07 | Improved |
| MOCHI TTO instead of classic registration pipeline | Classic Registrations · FaMoS test | median point-to-surface error (mm): 0.10 -> 0.07 | Improved |
| increase TTO iterations from 20 to 100 (ablation) | MOCHI TTO · FaMoS val | median point-to-surface error (mm): 0.08 -> 0.06 | Improved |
| use point-map and normal losses instead of point-to-surface loss in TTO (ablation) | MOCHI TTO · FaMoS val | median point-to-surface error (mm): 0.10 -> 0.08 at t=20; 0.06 -> 0.06 at t=100 | Mixed |
| use point-map losses instead of point-to-surface in coarse stage training (ablation) | MOCHI coarse stage · FaMoS | visual artifacts / self-intersections | Improved |
| add normal-map loss in TTO (ablation) | MOCHI TTO · FaMoS val | median point-to-surface error (mm): 0.09 -> 0.08 | Improved |
| add pretrained weights before TTO (vs optimizing from random initialization) (ablation) | MOCHI TTO · FaMoS val | median point-to-surface error (mm): 0.44 -> 0.07 | Improved |
| optimize refinement module weights instead of vertices directly (ablation) | MOCHI TTO · FaMoS val | median point-to-surface error (mm): 0.09 -> 0.07 | Improved |
| add PLIKS inverse parameter recovery on top of 2D landmarks (ablation) | MOCHI · FaMoS | mesh topology quality (qualitative) | Improved |
| add dense 2D landmark supervision to point-map loss plus PLIKS (ablation) | MOCHI · FaMoS | lip and eye alignment (qualitative) | Improved |
Closest papers by meaning
- Instant Multi-View Head Capture Through Learnable Registration (CVPR 2023 · similarity 0.90)
- Dense Semantic and Topological Correspondence of 3D Faces without Landmarks (ECCV 2018 · similarity 0.88)
- Shape My Face: Registering 3D Face Scans by Surface-to-Surface Translation (IJCV 2021 · similarity 0.88)
- Large Pose 3D Face Reconstruction From a Single Image via Direct Volumetric CNN Regression (ICCV 2017 · similarity 0.87)
- Self-Supervised Monocular 3D Face Reconstruction by Occlusion-Aware Multi-view Geometry Consistency (ECCV 2020 · similarity 0.87)
- A Sparse and Locally Coherent Morphable Face Model for Dense Semantic Correspondence Across Heterogeneous 3D Faces (TPAMI 2021 · 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.
