Research map / Point clouds / Registration and matching
SG-PGM: Partial Graph Matching Network with Semantic Geometric Fusion for 3D Scene Graph Alignment and Its Downstream Tasks
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 SG-PGM, a GNN-based partial graph matching network that fuses scene-graph semantics with point-cloud geometric features to align 3D scene graphs, and uses the alignment to support overlap checking, registration, and mosaicking.
Problem
Aligning 3D scene graphs between scene fragments (partial, low-overlap, with scene changes) as a first step for registration, mosaicking, and overlap checking, where prior SGAligner ignores explicit partial matching and reuses a weaker geometric encoder.
Approach
Formulates 3D scene graph alignment as partial graph matching solved by a GATv2-based network with Sinkhorn and a learnable soft top-k; fuses point-level KPConv geometric features from a pretrained GeoTransformer into node embeddings (P2SG); and rescores point correspondences in registration using the predicted node-level alignment matrix.
Key results
- On node matching with random transformations, SG-PGM keeps high accuracy while SGAligner drops significantly; on the T=I4 overlap check, SG-PGM@3 reaches 95.21 F1 versus 91.48 for SGAligner.
- On point cloud registration with T=I4, SG-PGM reduces RRE from 1.012 (SGAligner) to 0.6252 and RTE from 1.67 to 1.32.
- Point cloud mosaicking F1 reaches 99.80 for SG-PGM versus 93.58 for SGAligner.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add P2SG fusion of KP-Conv geometric features to the baseline graph matcher (ablation) | SG-PGM node matching · 3RScan | F1: 62.9 -> 79.0 | Helped |
| Replace KP-Conv features with PointNet features (ablation) | SG-PGM node matching · 3RScan | F1: 79.0 -> 79.1 | No clear effect |
| Add Soft-topK and AFA-U partial matching (ablation) | B+P node matching · 3RScan | F1: 79.0 -> 81.8 | Helped |
| Use AIS instead of simple matrix-product node similarity (ablation) | B+P+K node matching · 3RScan | F1: 81.8 -> 89.4 | Helped |
| Use Super-point Matching Rescoring to apply scene-graph alignment priors to point matches (ablation) | SG-PGM registration · 3RScan point cloud registration, random transformation, 10-30% overlap | SCC: 81.85 -> 87.60 | Helped |
| Use Graph-Cut RANSAC for pose estimation (ablation) | SG-PGM registration · 3RScan point cloud registration, T = I4 | RRE: 0.6252 -> 0.5103 | Helped |
| Use the top 3 matching scores instead of all matching scores for overlap scoring (ablation) | SG-PGM overlap checking · 3RScan | F1: 93.29 -> 95.21 | Helped |
| Register only overlap-to-overlap regions rather than all-to-all (ablation) | GeoTransformer with ground-truth scene-graph alignment · 3RScan point cloud registration | RRE: 1.9398 -> 1.4443 | Helped |
| Register object pairs one by one rather than all-to-all (ablation) | GeoTransformer with ground-truth scene-graph alignment · 3RScan point cloud registration | RRE: 1.9398 -> 5.9528 | Hurt |
Surprising findings
- Low-overlap cases (10-30%) show a notably higher false-positive rate and an F1 more than 20% lower than high-overlap cases.
- Registration on object-per-object fashion performed worst, while masking to the overlap region performed best.
Limitations
- More false-positive matches in low-overlap scenarios (F1 drops over 20% in 10-30% overlap range).
- Relies on a pretrained GeoTransformer encoder whose weights are locked, so geometric features are not adapted.
Experimental setup
Datasets: 3RScan 3DMatch
Benchmarks: 3RScan
Metrics: MRR Hits@1 Hits@3 Hits@5 F1 precision recall RRE RTE FMR
Baselines: SGAligner GeoTransformer SGA* B+P (PointNet variant) GCNet
Builds on: SGAligner GeoTransformer AFA-U / soft top-k GATv2 SimGNN
Closest papers by meaning
- Robust Point Cloud Registration Framework Based on Deep Graph Matching (CVPR 2021 · similarity 0.89)
- GraphI2P: Image-to-Point Cloud Registration with Exploring Pattern of Correspondence via Graph Learning (CVPR 2025 · similarity 0.88)
- SAG-GNN: Semantic-Aware Guided GNN for Descriptor-Free 2D-3D Matching (CVPR 2026 · similarity 0.87)
- Learning 3D Semantic Scene Graphs From 3D Indoor Reconstructions (CVPR 2020 · similarity 0.87)
- 3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration (NeurIPS 2024 · similarity 0.86)
- DGC-GNN: Leveraging Geometry and Color Cues for Visual Descriptor-Free 2D-3D Matching (CVPR 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.