Research map / Point clouds / Shape and surface completion
PFCNN: Convolutional Neural Networks on 3D Surfaces Using Parallel Frames
Check 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 PFCNN instead of MDGCNN | human body segmentation on original meshes · Human body segmentation (original meshes) | accuracy (%): 88.2 -> 91.45 | Helped |
| Use PFCNN instead of MDGCNN | human body segmentation on remeshed meshes · Human body segmentation (remeshed meshes) | accuracy (%): 89.53 -> 91.79 | Helped |
| Use PFCNN instead of TextureNet | ScanNet validation segmentation · ScanNet validation | mIoU: 0.58 -> 0.632 | Helped |
| Increase the PFCNN network size (ablation) | ScanNet validation segmentation · ScanNet validation | mIoU: 0.632 -> 0.662 | Helped |
| Use principal curvature frames as a 4-direction frame field (ablation) | PFCNN registration · FAUST non-rigid registration (vertex classification) | accuracy (%): 83.29 -> 89.80 | Helped |
| Optimize a parallel frame field aligned to geometric features (ablation) | PFCNN registration · FAUST non-rigid registration (vertex classification) | accuracy (%): 89.80 -> 92.01 | Helped |
| Increase frame-field symmetry order from N=1 to N=4 (ablation) | PFCNN registration · FAUST non-rigid registration (vertex classification) | accuracy (%): 83.11 -> 92.01 | Helped |
| Increase frame-field symmetry order from N=4 to N=8 (ablation) | PFCNN registration · FAUST non-rigid registration (vertex classification) | accuracy (%): 92.01 -> 93.35 | Helped |
| Increase frame-field symmetry order from N=4 to N=8 (ablation) | PFCNN registration · FAUST non-rigid registration (vertex classification) | runtime (ms): 139.10 -> 227.72 | Hurt |
| Use instance normalization instead of batch normalization (ablation) | PFCNN shape classification · SHREC'15 non-rigid shape classification | accuracy (%): 11 -> 99.5 | Helped |
| Use mild curvature-direction alignment (λ=0.01) instead of smoothness-only frames (ablation) | PFCNN registration · FAUST non-rigid registration (vertex classification) | accuracy (%): 88.56 -> 92.01 | Helped |
| Increase curvature-direction alignment weight from λ=0.01 to λ=0.1 or λ=1 (ablation) | PFCNN registration · FAUST non-rigid registration (vertex classification) | accuracy (%): 92.01 -> 91.97; 92.01 -> 90.77 | Hurt |
Closest papers by meaning
- TextureNet: Consistent Local Parametrizations for Learning From High-Resolution Signals on Meshes (CVPR 2019 · similarity 0.88)
- SurfNet: Generating 3D Shape Surfaces Using Deep Residual Networks (CVPR 2017 · similarity 0.87)
- DualConvMesh-Net: Joint Geodesic and Euclidean Convolutions on 3D Meshes (CVPR 2020 · similarity 0.87)
- Primal-Dual Mesh Convolutional Neural Networks (NeurIPS 2020 · similarity 0.87)
- Flatten Anything: Unsupervised Neural Surface Parameterization (NeurIPS 2024 · similarity 0.85)
- Improving Neural Network Surface Processing with Principal Curvatures (NeurIPS 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.