Research map / 3D vision / Face and shape
Parallelised Differentiable Straightest Geodesics for 3D Meshes
Hippolyte Verninas, Caner Korkmaz, Stefanos Zafeiriou, Tolga Birdal, Simone Foti
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 GPU-parallel straightest geodesics instead of CPU implementations | Exponential map · large batch sizes | runtime: 3 orders of magnitude faster | Improved |
| use GFD instead of EP differentiation (ablation) | Exp backward pass · sphere gradient check | gradient accuracy: backward about 3.85 times more expensive than EP | Mixed |
| keep gradients w.r.t. start point p in EP (ablation) | Extrinsic proxy · sphere gradient check | gradient accuracy | Hurt |
| use our Exp instead of projection integration (PI) | Exponential map · high face count meshes | runtime: 5 to 6 orders of magnitude faster | Improved |
| use learned patch size via differentiable Exp (AGC) instead of fixed-patch GCNN | GCNN · body part segmentation | accuracy: 86.4% -> 91.7% (XYZ), 92.3% (HKS) | Improved |
| use HKS input instead of XYZ coordinates (ablation) | AGC · body part segmentation | accuracy: 91.7% -> 92.3% | Improved |
| use GFD instead of EP for AGC training (ablation) | AGC · body part segmentation | accuracy and training time: 1 min 21 s -> 2 min 8 s per epoch | No clear effect |
| use MeshFlow with exponential map instead of RFM | Riemannian flow matching on meshes · Bunny/Spot eigenfunctions | inference speed and memory: 1.6e4 times faster, 97% less GPU memory | Improved |
| use MeshFlow instead of RFM | Riemannian flow matching on meshes · Bunny/Spot eigenfunctions | NLL: 1.55 -> 1.71 (Bunny k=100) | Hurt |
| use more samples (1024 vs 256) for OT coupling (ablation) | MeshFlow · Bunny k=100 | KLD: 0.13 -> 0.058 | Improved |
| use Mesh-LBFGS instead of Lloyd's algorithm | Geodesic centroidal Voronoi tessellation · Spot the Cow, 50 seeds | convergence and final energy | Improved |
| start seeds clustered rather than uniform (ablation) | Mesh-LBFGS · Spot the Cow, 50 seeds | advantage over Lloyd | Improved |
Closest papers by meaning
- Parallel and Scalable Heat Methods for Geodesic Distance Computation (TPAMI 2019 · similarity 0.89)
- Neural Surface Maps (CVPR 2021 · similarity 0.87)
- Flatten Anything: Unsupervised Neural Surface Parameterization (NeurIPS 2024 · similarity 0.86)
- Flow Matching on General Geometries (ICLR 2024 · similarity 0.85)
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 2024 · similarity 0.85)
- Dynamic Neural Surfaces for Elastic 4D Shape Representation and Analysis (CVPR 2025 · similarity 0.84)
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.
