Research map / 3D vision / 3D generation
Symmetry Strikes Back: From Single-Image Symmetry Detection to 3D Generation
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 Reflect3D, a transformer-based zero-shot detector of 3D reflection symmetry from a single RGB image, aided by multi-view diffusion views, and uses detected symmetry to improve DreamGaussian-based single-image 3D generation.
Problem
Detecting 3D reflection symmetry planes from a single RGB image of arbitrary objects in a zero-shot manner, and exploiting that symmetry to improve single-image 3D generation.
Approach
A feed-forward transformer with a frozen DINOv2 encoder queries hypothesis normals via cross-attention, classifying and regressing symmetry planes; multi-view diffusion synthesizes surrounding views whose predictions are CLIP-filtered, rotated back, and K-Means clustered to resolve single-view ambiguity. Detected symmetry is then injected into DreamGaussian via symmetry alignment, symmetric SDS/de
Key results
- On GSO, Reflect3D reaches F@5 of 0.390 versus 0.040 for NeRD, and geodesic distance 13.3 versus 36.2.
- On OmniObject3D, Reflect3D achieves F@5 of 0.173 versus 0.055 for NeRD and GD of 22.8 versus 41.3.
- Symmetry-aware generation improves DreamGaussian on GSO with CLIP-Sim 0.629 vs 0.592 and Chamfer distance 0.414 vs 0.442.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multi-view diffusion and clustering-based aggregation | Reflect3D-FF · GSO | F@5°: 0.191 -> 0.390 | Helped |
| Add multi-view diffusion and clustering-based aggregation | Reflect3D-FF · OmniObject | F@5°: 0.103 -> 0.173 | Helped |
| Add symmetry guidance, alignment, symmetric SDS, densification, and texture refinement | DreamGaussian · GSO | F@0.2: 0.404 -> 0.451 | Helped |
| Add symmetry guidance, alignment, symmetric SDS, densification, and texture refinement | DreamGaussian · OmniObject | F@0.2: 0.427 -> 0.457 | Helped |
| Remove clustering-based multi-view aggregation (ablation) | Reflect3D · GSO | F@15°: 0.756 -> 0.655 | Hurt |
| Replace DINOv2 with ImageNet-trained ViT (ablation) | Reflect3D · GSO | F@15°: 0.756 -> 0.386 | Hurt |
| Fine-tune DINOv2 instead of freezing it (ablation) | Reflect3D · GSO | F@15°: 0.756 -> 0.236 | Hurt |
| Remove symmetry alignment (ablation) | Symmetry-aware DreamGaussian · Single-image 3D generation | geometry and texture quality | Hurt |
| Remove symmetric SDS (ablation) | Symmetry-aware DreamGaussian · Single-image 3D generation | geometry and symmetry quality | Hurt |
| Remove symmetric densification (ablation) | Symmetry-aware DreamGaussian · Single-image 3D generation | geometry quality | Hurt |
| Remove symmetric texture refinement (ablation) | Symmetry-aware DreamGaussian · Single-image 3D generation | texture quality | Hurt |
Surprising findings
- Fine-tuning the DINOv2 encoder drastically degrades symmetry detection (F@5 drops from 0.390 to 0.038 on GSO), supporting the frozen-feature design.
Limitations
- Not designed for completely asymmetric or highly deformable objects.
- Symmetry plane absolute distance is not resolved from single view and relies on generated 3D representation.
- Training symmetry ground truth is generated by an optimization procedure and the exact training datasets are not released in full for some baselines.
Experimental setup
Datasets: Objaverse ShapeNet GSO OmniObject3D
Benchmarks: GSO OmniObject3D
Metrics: F@5 F@15 F@30 F@50 geodesic distance CLIP similarity Chamfer distance F-score@0.1/0.2/0.5
Baselines: NeRD Shape to Symmetry DreamGaussian Random Guess
Builds on: NeRD DreamGaussian Zero-1-to-3 DINOv2
Closest papers by meaning
- Leveraging 3D Geometric Priors in 2D Rotation Symmetry Detection (CVPR 2025 · similarity 0.87)
- GIQ: Benchmarking 3D Geometric Reasoning of Vision Foundation Models with Simulated and Real Polyhedra (ICLR 2026 · similarity 0.86)
- HarmonyView: Harmonizing Consistency and Diversity in One-Image-to-3D (CVPR 2024 · similarity 0.86)
- Pygmalion Effect in Vision: Image-to-Clay Translation for Reflective Geometry Reconstruction (NeurIPS 2026 · similarity 0.86)
- Align3D-AD: Cross-Modal Feature Alignment and Dual-Prompt Learning for Zero-shot 3D Anomaly Detection (NeurIPS 2026 · similarity 0.85)
- SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation (NeurIPS 2026 · 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.