Research map / 3D vision / Face and shape
ZeroShape: Regression-based Zero-shot Shape Reconstruction
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 ZeroShape, a regression-based single-image zero-shot 3D shape reconstruction model that estimates depth and camera intrinsics to unproject a visible-surface projection map, then predicts occupancy, and curates a large real-world benchmark.
Problem
Single-image zero-shot 3D shape reconstruction; generative methods generalize but are computationally expensive, and prior evaluations are small and inconsistent.
Approach
A regression model that predicts depth and camera intrinsics jointly, unprojects them into a normalized visible-surface projection map, and uses cross-attention over its patch features to predict occupancy, trained only on synthetic renders.
Key results
- On OmniObject3D, ZeroShape reaches FS@1 0.2297 and CD 0.310 versus 0.1683 and 0.407 for the best baseline OpenLRM.
- On Ocrtoc3D, ZeroShape reaches FS@1 0.2410 and CD 0.286, outperforming MCC (0.1994, 0.411).
- On Pix3D, ZeroShape (FS@1 0.1928) trails Shap-E (0.2016) on FS@1 but is competitive on FS@2 (0.4290 vs 0.4287) and FS@5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ZeroShape instead of OpenLRM | OmniObject3D shape reconstruction · OmniObject3D | FS@5: 0.7204 -> 0.8169 | Helped |
| use ZeroShape instead of Shap-E | Pix3D shape reconstruction · Pix3D | FS@5: 0.7833 -> 0.7759 | Hurt |
| use ZeroShape instead of MCC | Ocrtoc3D shape reconstruction · Ocrtoc3D | FS@5: 0.7135 -> 0.8459 | Helped |
| add explicit geometric reasoning with depth and camera pretraining losses (ablation) | ZeroShape · OmniObject3D | FS@5: 0.7797 -> 0.8169 | Helped |
| use unprojection to form a 3D projection map instead of depth-map features (ablation) | ZeroShape · OmniObject3D | FS@5: 0.8053 -> 0.8169 | Helped |
| learn camera intrinsics instead of using fixed intrinsics (ablation) | ZeroShape · OmniObject3D | FS@5: 0.8039 -> 0.8169 | Helped |
Surprising findings
- On Pix3D, Shap-E and Point-E outperform ZeroShape on several metrics, which the authors attribute to furniture being abundant in those models' training sets.
Limitations
- Pix3D results are dominated by furniture categories, so the benchmark is biased toward chairs and sofas.
- Trained only on synthetic renders, and the benchmark is built from existing datasets, with some images rendered from 3D scans rather than real photos.
Experimental setup
Datasets: ShapeNetCore.v2 Objaverse-LVIS OmniObject3D Ocrtoc3D Pix3D
Benchmarks: OmniObject3D Ocrtoc3D Pix3D
Metrics: F-score Chamfer Distance
Baselines: SS3D MCC One-2-3-45 OpenLRM Point-E Shap-E
Builds on: MCC DPT Omnidata Zero-123
Closest papers by meaning
- Robust 3D Shape Reconstruction in Zero-Shot from a Single Image in the Wild (CVPR 2025 · similarity 0.90)
- LRM-Zero: Training Large Reconstruction Models with Synthesized Data (NeurIPS 2024 · similarity 0.88)
- GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors (CVPR 2025 · similarity 0.86)
- Towards In-the-wild 3D Plane Reconstruction from a Single Image (CVPR 2025 · similarity 0.86)
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.86)
- ZeroNVS: Zero-Shot 360-Degree View Synthesis from a Single Image (CVPR 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.