Research map / 3D vision / 3D generation
One-shot 3D Object Canonicalization based on Geometric and Semantic Consistency
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 a one-shot category-level 3D object canonicalization framework that uses one prior model per category, LLM/VLM zero-shot part perception, a support-plane multi-hypothesis initialization, and a joint geometric-semantic energy, and releases the Canonical Objaverse Dataset (32K shapes, 1,054 categories).
Problem
3D object canonicalization for arbitrary categories is hard for learning-based methods that need many per-category training samples, which real long-tailed datasets like Objaverse lack.
Approach
Use a single prior model per category, zero-shot semantic part correspondences from ChatGPT and GLIP, a support-plane strategy to generate initial poses, a joint energy combining a semantic Gaussian overlap term with a Chamfer-distance Gaussian mixture, and semantic relation maps to select the final canonical pose.
Key results
- With one prior model, average IC/GEC on ShapeNet is 0.703/0.784 for the proposed method versus 1.401/2.162 for ConDor (best baseline with 10 priors), and the paper reports at least 49.8% and 63.1% average IC and GEC improvements.
- On DREDS, the method reaches average IC/GEC of 0.194/0.208 versus 0.806/1.158 for ConDor.
- Canonicalizes 32K Objaverse-LVIS shapes into the Canonical Objaverse Dataset spanning 1,054 categories.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use one-shot canonicalization instead of few-shot baselines | Ours · ShapeNet (Avg.) | GEC: 2.123 -> 0.784 | Helped |
| use one-shot canonicalization instead of ConDor | Ours · NOCS (Avg.) | GEC: 0.876 -> 0.129 | Helped |
| use one-shot canonicalization instead of ConDor | Ours · DREDS (Avg.) | GEC: 1.158 -> 0.208 | Helped |
| add semantic guidance to geometric alignment rather than use the geometric constraint alone (ablation) | Ours · DREDS | GEC: 0.724 -> 0.208 | Helped |
| add geometric guidance to semantic alignment rather than use the semantic constraint alone (ablation) | Ours · DREDS | GEC: 2.315 -> 0.208 | Helped |
| use the joint energy function instead of a naive weighted combination of geometric and semantic constraints (ablation) | Ours · DREDS | GEC: 0.870 -> 0.208 | Helped |
| add multi-hypothesis pose initialization and semantic-map selection (ablation) | Ours · DREDS | GEC: 0.690 -> 0.208 | Helped |
Surprising findings
- Naive weighted combination of geometric and semantic constraints performs worse than geometric constraints alone on DREDS (IC 0.802 vs 0.696).
Limitations
- Relies on zero-shot GLIP detections that degrade on uncommon poses, which the support-plane strategy only partially mitigates.
- Evaluated on relatively few categories and small real datasets; the wild-data evaluation is qualitative only.
- Requires a single prior model per category, so prior quality and choice may affect results.
Experimental setup
Datasets: ShapeNet NOCS DREDS OmniObject3D Objaverse-LVIS
Benchmarks: ShapeNet NOCS DREDS OmniObject3D Objaverse-LVIS
Metrics: Instance-Level Consistency (IC) Ground Truth Equivariance Consistency (GEC)
Baselines: CaCa ConDor ShapeMatcher
Builds on: ConDor CaCa ShapeMatcher GLIP
Closest papers by meaning
- Emergence of a Shared Canonical Object Frame from In-the-Wild Videos (NeurIPS 2026 · similarity 0.88)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.88)
- Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.87)
- 3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation (ICLR 2024 · similarity 0.87)
- Orientation Matters: Making 3D Generative Models Orientation-Aligned (NeurIPS 2025 · similarity 0.87)
- ShapeMatcher: Self-Supervised Joint Shape Canonicalization Segmentation Retrieval and Deformation (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.