Research map / Point clouds / Shape and surface completion
OffsetOPT: Explicit Surface Reconstruction without Normals
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains a transformer that predicts surface triangles from local KNN geometry on uniformly sampled ABC meshes, then optimizes per-point offsets with the frozen network to reconstruct explicit surfaces from normal-free point clouds.
Problem
Explicit surface reconstruction from raw 3D point clouds without normals, producing edge-manifold triangles and preserving sharp features, where prior neural computational methods are restricted to Poisson-disk sampled inputs and implicit methods need normals.
Approach
Instead of requiring ideally distributed inputs, a frozen triangle-prediction network is adapted to arbitrary point clouds by optimizing per-point offsets through unsupervised BCE loss on the network's own triangle predictions, with controlled updates that prevent points from colliding or drifting off the surface.
Key results
- On FAUST, OffsetOPT achieves CD1 0.217 and EF1 0.896 versus 0.218/0.801 for DSE and 0.302/0.501 for NKSR (+n).
- On ScanNet, OffsetOPT reaches CD1 0.147 and ECD1 0.389 (NR), outperforming NKSR (+n) at 0.157 CD1 without normals required.
- Offset optimization raises manifold-edge percentage from 82% to 98% on FAUST and from 74% to 99% on ScanNet (Table 5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add offset optimization (ablation) | OffsetOPT · FAUST | CD1 (×102): 0.221 -> 0.217 | Helped |
| add offset optimization (ablation) | OffsetOPT · MGN | CD1 (×102): 0.280 -> 0.278 | Helped |
| add offset optimization (ablation) | OffsetOPT · ScanNet | CD1 (×102): 0.154 -> 0.147 | Helped |
| add offset optimization (ablation) | OffsetOPT · Matterport3D | CD1 (×102): 0.164 -> 0.148 | Helped |
| use the proposed nonzero offset initialization instead of zero initialization (ablation) | OffsetOPT · FAUST and Thingi10k | percentage of manifold edges | Helped |
| use offset optimization | OffsetOPT · MGN | CD1 (×102): 0.269 -> 0.278 | Hurt |
| use offset optimization | OffsetOPT · MGN | ECD1 (×102): 3.231 -> 5.447 | Hurt |
Surprising findings
- Without ground-truth normals, OffsetOPT outperforms NKSR given ground-truth normals on large-scale scenes in several metrics.
Limitations
- A small percentage of non-manifold edges may remain, requiring optional post-processing.
- Transformer architecture is less efficient than graph convolutional networks used in CircNet.
- Scene reconstruction takes minutes per scene; method assumes input points adequately capture the underlying surface (noisy/sparse inputs out of scope).
Experimental setup
Datasets: ABC FAUST MGN ScanNet Matterport3D CARLA Thingi10k
Benchmarks: ABC FAUST MGN ScanNet Matterport3D CARLA
Metrics: CD1 CD2 F1 NC NR ECD1 EF1 manifold edge percentage
Baselines: ball-pivoting PointTriNet DSE CircNet SPSR NKSR
Builds on: CircNet NKSR PointTriNet DSE
Closest papers by meaning
- Reconstructing Surfaces for Sparse Point Clouds With On-Surface Priors (CVPR 2022 · similarity 0.88)
- Deep Geometric Prior for Surface Reconstruction (CVPR 2019 · similarity 0.88)
- Iso-Points: Optimizing Neural Implicit Surfaces With Hybrid Representations (CVPR 2021 · similarity 0.88)
- Sign-Agnostic Implicit Learning of Surface Self-Similarities for Shape Modeling and Reconstruction From Raw Point Clouds (CVPR 2021 · similarity 0.88)
- POCO: Point Convolution for Surface Reconstruction (CVPR 2022 · similarity 0.87)
- GridFormer: Point-Grid Transformer for Surface Reconstruction (AAAI 2024 · similarity 0.86)
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.