Research map / 3D vision / Face and shape
EchoMatch: Partial-to-Partial Shape Matching via Correspondence Reflection
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 EchoMatch, a partial-to-partial 3D shape matching framework that predicts overlap via forward-backward correspondence reflection inside a functional map pipeline, using DINOv2 image features.
Problem
Partial-to-partial 3D shape matching requires both finding correspondences and determining which points of each shape have a partner; existing methods struggle with overlap prediction and rely on spatial coordinate features that depend on shape alignment.
Approach
Model matching in each direction separately with row-stochastic soft point maps, then compose X→Y and Y→X; points that return to their original neighbourhood are overlapping, others are not. The reflection scores are refined by a DiffusionNet and used within a functional map framework with DINOv2 image features.
Key results
- EchoMatch with DINOv2 features reaches mIoU 84.72 on CP2P24 and 84.75 on PSMAL, versus 74.15 and 73.67 for DPFM.
- On BeCoS with DINOv2 features, mIoU is 64.68 versus 51.02 for DPFM.
- Ablation: neighbourhood size 64 gives the best BeCoS validation mIoU (67.43 vs 55.04 at size 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DINOv2 features instead of XYZ input features | EchoMatch · CP2P24 | mIoU (×100): 80.10 -> 84.72 | Helped |
| Use DINOv2 features instead of XYZ input features | EchoMatch · PSMAL | mIoU (×100): 72.71 -> 84.75 | Helped |
| Use DINOv2 features instead of XYZ input features | EchoMatch · BeCoS | mIoU (×100): 52.40 -> 64.68 | Helped |
| Use EchoMatch instead of DPFM with DINOv2 features | partial-to-partial shape matching · CP2P24 | AUC: 72.85 -> 84.61 | Helped |
| Use EchoMatch instead of DPFM with DINOv2 features | partial-to-partial shape matching · PSMAL | AUC: 67.13 -> 81.27 | Helped |
| Use EchoMatch instead of DPFM with DINOv2 features | partial-to-partial shape matching · BeCoS | AUC: 45.96 -> 59.18 | Helped |
| Increase correspondence-reflection neighbourhood size from 1 to 64 (ablation) | EchoMatch · BeCoS validation | mIoU (×100): 55.04 -> 67.43 | Helped |
| Increase correspondence-reflection neighbourhood size from 1 to 64 (ablation) | EchoMatch · BeCoS validation | Geodesic Error (×100): 7.96 -> 7.80 | Helped |
| Prune FY to the ground-truth overlap before functional-map computation (ablation) | EchoMatch · BeCoS validation | Geodesic Error (×100): 8.72 -> 7.80 | Helped |
| Prune features on both shapes instead of pruning FY only (ablation) | EchoMatch · BeCoS validation | mIoU (×100): 67.43 -> 65.48 | Hurt |
| Prune features on both shapes instead of pruning FY only (ablation) | EchoMatch · BeCoS validation | Geodesic Error (×100): 7.80 -> 8.80 | Hurt |
| Train with DINOv2 rather than spatial coordinate features (ablation) | EchoMatch · CP2P21 randomly rotated test set | average geodesic error: 12.36 -> 9.73 | Helped |
Surprising findings
- With spatial XYZ coordinates on BeCoS, all methods perform near random, so XYZ-based comparisons there are not meaningful.
Limitations
- Depends on quality of 2D image features; points not visible in any image must be approximated.
- Image feature quality is lower for partial shapes than for complete shapes.
- Not fully invariant to mesh discretisation.
Experimental setup
Datasets: CP2P24 PARTIALSMAL BeCoS
Benchmarks: CP2P24 PSMAL BeCoS
Metrics: mean IoU geodesic error PCK curves AUC
Baselines: DPFM SM-COMB GC-PPSM
Builds on: DPFM functional maps Diff3f DiffusionNet PointInfoNCE
Closest papers by meaning
- Partial-to-Partial Shape Matching with Geometric Consistency (CVPR 2024 · similarity 0.92)
- Multi-Shape Matching with Cycle Consistency Basis via Functional Maps (AAAI 2025 · similarity 0.86)
- Wormhole Loss for Partial Shape Matching (NeurIPS 2024 · similarity 0.86)
- Stable-SCore: A Stable Registration-based Framework for 3D Shape Correspondence (CVPR 2025 · similarity 0.85)
- Shape Abstraction via Marching Differentiable Support Functions (CVPR 2025 · similarity 0.85)
- 3D Geometric Shape Assembly via Efficient Point Cloud Matching (ICML 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.