Research map / Point clouds / Shape and surface completion
PatchNets: Patch-Based Generalizable Deep Implicit 3D Shape Representations
Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Carsten Stoll, Christian Theobalt
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 patch-based representation instead of single global patch | DeepSDF-style baseline · ShapeNet (100 shapes per category) | IoU: 76.5 -> 92.1 | Improved |
| use patch-based representation instead of single global patch | DeepSDF-style baseline · ShapeNet (100 shapes per category) | F-score: 80.6 -> 94.8 | Improved |
| train on only Cabinet category: patches instead of global patch | global-patch baseline · ShapeNet (trained on Cabinet) | IoU: 66.4 -> 90.3 | Improved |
| cross-dataset test with patches instead of global patch | global-patch baseline · Dynamic FAUST (trained on Cabinet) | F-score: 25.0 -> 94.0 | Improved |
| increase number of patches from 3 to default (ablation) | PatchNets · ShapeNet reduced test set | IoU: 73.8 -> 91.6 | Improved |
| increase number of patches from 10 to default (ablation) | PatchNets · ShapeNet reduced test set | IoU: 85.2 -> 91.6 | Improved |
| increase patch latent/hidden size from 32 to default (ablation) | PatchNets · ShapeNet reduced test set | IoU: 82.8 -> 91.6 | Improved |
| increase latent/hidden size to 512 (ablation) | PatchNets · ShapeNet reduced test set | IoU: 91.6 -> 95.3 | Improved |
| train on full dataset instead of 100 shapes per category (ablation) | PatchNets · ShapeNet reduced test set | IoU: 91.6 -> 92.2 | Improved |
| add coverage loss L_cov (ablation) | PatchNets · ShapeNet reduced test set | IoU: 90.7 -> 91.6 | Improved |
| add rotation loss L_rot (ablation) | PatchNets · ShapeNet reduced test set | IoU: 92.5 -> 91.6 | Hurt |
| use normal-aligned rotation initialization with L_rot instead of random initialization (ablation) | PatchNets · ShapeNet reduced test set | IoU: 89.0 -> 91.6 | Improved |
Closest papers by meaning
- DeepCurrents: Learning Implicit Representations of Shapes With Boundaries (CVPR 2022 · similarity 0.91)
- Implicit Surface Representations As Layers in Neural Networks (ICCV 2019 · similarity 0.91)
- PatchComplete: Learning Multi-Resolution Patch Priors for 3D Shape Completion on Unseen Categories (NeurIPS 2022 · similarity 0.91)
- Learning Manifold Patch-Based Representations of Man-Made Shapes (ICLR 2021 · similarity 0.91)
- Learning to Infer Implicit Surfaces without 3D Supervision (NeurIPS 2019 · similarity 0.90)
- Learning 3D Equivariant Implicit Function with Patch-Level Pose-Invariant Representation (NeurIPS 2024 · similarity 0.88)
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.
