Research map / 3D vision / Face and shape
GenCorres: Consistent Shape Matching via Coupled Implicit-Explicit Shape Generative Models
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.
Learns an implicit shape generator with geometric (ARAP/ACAP) and cycle-consistency regularizers, then converts it to a mesh generator to get consistent correspondences for unsupervised joint shape matching.
Problem
Unsupervised joint shape matching (JSM) over deformable shape collections, where pairwise matching fails on dissimilar shapes and consistent dense correspondences are costly to store for large collections.
Approach
Induce dense correspondences between adjacent implicit surfaces via a constrained QP over ARAP/ACAP energies in latent space, use these as regularizers (geometric deformation and 3-cycle consistency) for an implicit VAE generator, then fit an explicit template-based mesh generator whose outputs give consistent correspondences.
Key results
- GenCorres reduces JSM mean/median geodesic error by 62.0%/66.8% on DFAUST vs. the best JSM baseline (MIM/CZO), reaching 1.30/1.13 cm.
- On shape generation, GenCorres reaches 1.83 cm mean reconstruction error on DFAUST vs. 1.88 for SALD and 6.85 cm on SMAL vs. 7.66 for SALD.
- On FAUST pairwise matching, GenCorres achieves 1.6 cm error, beating template-based 3D-CODED (2.5) and IFS (2.6) and matching spectral methods such as ULRSSM (1.6).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GenCorres instead of SALD | shape generation · DFAUST | mean reconstruction error (cm): 1.88 -> 1.83 | Helped |
| use GenCorres instead of SALD | shape generation · SMAL | mean reconstruction error (cm): 7.66 -> 6.85 | Helped |
| use GenCorres instead of NeuroMorph | joint shape matching · DFAUST | mean geodesic error (cm): 2.49 -> 1.30 | Helped |
| use GenCorres instead of CZO | joint shape matching · SMAL | median geodesic error (cm): 1.12 -> 0.47 | Helped |
| add cycle-consistency regularization (ablation) | GenCorres · DFAUST | mean geodesic error (cm): 1.41 -> 1.30 | Helped |
| add cycle-consistency regularization (ablation) | GenCorres · SMAL | mean geodesic error (cm): 1.11 -> 1.02 | Helped |
| add geometric deformation regularization (ablation) | GenCorres · DFAUST | mean geodesic error (cm): 7.65 -> 1.30 | Helped |
| add geometric deformation regularization (ablation) | GenCorres · SMAL | mean geodesic error (cm): 6.28 -> 1.02 | Helped |
| use ACAP rather than ARAP regularization (ablation) | GenCorres · DFAUST | mean geodesic error (cm): 1.62 -> 1.30 | Helped |
| use ACAP rather than ARAP regularization (ablation) | GenCorres · SMAL | mean geodesic error (cm): 1.07 -> 1.02 | Helped |
| use the explicit mesh generator instead of propagated implicit-generator correspondences (ablation) | GenCorres · DFAUST | mean geodesic error (cm): 2.62 -> 1.30 | Helped |
| use the explicit mesh generator instead of propagated implicit-generator correspondences (ablation) | GenCorres · SMAL | mean geodesic error (cm): 1.24 -> 1.02 | Helped |
Surprising findings
- Removing the geometric deformation regularizer increases DFAUST mean geodesic error by 488%, showing that cycle consistency alone is insufficient without the ARAP/ACAP constraints.
Limitations
- Requires a reasonably large input shape collection; does not work with few input shapes, where pairwise matching is better.
- Regularizers depend on mesh discretization of implicit surfaces via Marching Cubes.
- Template-based approach does not use intrinsic features, so pairwise FAUST performance is only comparable to spectral methods.
Experimental setup
Datasets: DFAUST FAUST SMAL
Benchmarks: DFAUST SMAL FAUST
Metrics: Chamfer distance reconstruction error (cm) mean geodesic error (cm) median geodesic error (cm)
Baselines: LGF DPM DeepSDF SALD CZO MIM NeuroMorph 3D-CODED IFS BCICP ZO S-Shells
Builds on: SALD ARAPReg DeepSDF NeuroMorph
Closest papers by meaning
- Enhancing Implicit Shape Generators Using Topological Regularizations (ICML 2024 · similarity 0.88)
- Probabilistic Deformation Consistency for Unsupervised Shape Matching (AAAI 2026 · similarity 0.87)
- Spectral Meets Spatial: Harmonising 3D Shape Matching and Interpolation (CVPR 2024 · similarity 0.85)
- Flow-Corrected Shape Optimization: Taming Manifold Drift in High-Dimensional 3D Models (NeurIPS 2026 · similarity 0.85)
- HOPE: Shape Matching Via Aligning Different K-hop Neighbourhoods (NeurIPS 2024 · similarity 0.85)
- Wormhole Loss for Partial Shape Matching (NeurIPS 2024 · similarity 0.84)
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.