Research map / Point clouds / Shape and surface completion
GridFormer: Point-Grid Transformer for Surface Reconstruction
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.
Proposes GridFormer, a point-grid attention encoder that treats grid features as transfer points between space and point cloud, plus a two-stage margin BCE and boundary sampling training strategy for implicit surface reconstruction.
Problem
Implicit neural surface reconstruction from discrete point clouds: regular grid features lose shape detail while irregular point features are costly and hard to connect to continuous space, and occupancy functions are discontinuous at the surface, blurring boundaries.
Approach
Learn attention weights between the input points and a regular grid (plane/volume) so grid features carry point-level detail efficiently, with grid features as transfer points, and refine occupancy near the surface via a margin BCE loss and boundary sampling in a second training stage.
Key results
- On ShapeNet with 3000 points and sigma=0.005, GridFormer reaches IoU 0.936 and CD 0.28, versus ALTO at 0.930 and 0.30.
- On Synthetic Rooms, GridFormer achieves IoU 0.918 and CD 0.34, outperforming ALTO (0.914, 0.35) and POCO (0.884, 0.36).
- On ScanNet-v2 generalization (trained on Synthetic Rooms), GridFormer reaches CD 0.71 and FS 0.846, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use point-grid attention with a volume grid ($64^3$) instead of triplane grid ($3 * 128^2$) (ablation) | GridFormer · Synthetic Rooms | IoU: 0.835 -> 0.918 | Helped |
| remove grid downsampling (ablation) | GridFormer · ShapeNet | IoU: 0.942 -> 0.897 | Hurt |
| remove grid downsampling (ablation) | GridFormer · ShapeNet | Encode Time (s): 0.41 -> 0.39 | Helped |
| add boundary optimization (boundary sampling and margin binary cross-entropy) (ablation) | GridFormer · ShapeNet, 3000 Pts, σ = 0.005 | L2-CD (× 10^4): 0.211 -> 0.201 | Helped |
| add boundary optimization (boundary sampling and margin binary cross-entropy) (ablation) | GridFormer · ShapeNet, 3000 Pts, σ = 0.025 | L2-CD (× 10^4): 0.881 -> 0.718 | Helped |
| add boundary optimization (boundary sampling and margin binary cross-entropy) (ablation) | GridFormer · ShapeNet, 1000 Pts, σ = 0 | L2-CD (× 10^4): 0.253 -> 0.247 | Helped |
| add boundary optimization (boundary sampling and margin binary cross-entropy) (ablation) | GridFormer · ShapeNet, 1000 Pts, σ = 0.005 | L2-CD (× 10^4): 0.371 -> 0.328 | Helped |
| use GridFormer instead of ALTO | ShapeNet, 3000 Pts, σ = 0.005 · ShapeNet | IoU: 0.930 -> 0.936 | Helped |
| use GridFormer instead of ALTO | Synthetic Rooms · Synthetic Rooms | IoU: 0.914 -> 0.918 | Helped |
| use GridFormer instead of POCO | ScanNet-v2 without floors · ScanNet-v2 | CD: 0.74 -> 0.71 | Helped |
| use GridFormer rather than ConvONet | ShapeNet chairs · GridFormer and ConvONet | Inference Time (s): 0.43 -> 5.61 | Hurt |
Surprising findings
- The improvement over baselines shrinks with very sparse input (300 points), since learning per-point weights becomes less meaningful.
- Removing grid downsampling reduces encoding time slightly but significantly degrades reconstruction quality.
Limitations
- Performance gains diminish with very sparse point clouds (300 points).
- Results depend on grid resolution and point density.
- Trained on Synthetic Rooms, which contains regular walls and floors, so generalization to real scans completes irregular structures.
Experimental setup
Datasets: ShapeNet Synthetic Rooms ScanNet-v2
Benchmarks: ShapeNet Synthetic Rooms ScanNet-v2
Metrics: IoU Chamfer-L1 distance normal consistency F-Score L2-CD
Baselines: ONet ConvONet POCO ALTO SPSR SAP NKSR DP-ConvONet
Builds on: Point Transformer ConvONet ALTO Occupancy Networks MISE
Closest papers by meaning
- Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 2024 · similarity 0.86)
- OffsetOPT: Explicit Surface Reconstruction without Normals (CVPR 2025 · similarity 0.86)
- High-Fidelity Lightweight Mesh Reconstruction from Point Clouds (CVPR 2025 · similarity 0.86)
- Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation (CVPR 2025 · similarity 0.85)
- MeshFormer : High-Quality Mesh Generation with 3D-Guided Reconstruction Model (NeurIPS 2024 · similarity 0.85)
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · 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.