Research map / 3D vision / 3D generation
Slice3D: Multi-Slice Occlusion-Revealing Single View 3D 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 Slice3D, which predicts multi-slice images from a single RGB view (by regression with a U-Net or generation with a DDPM) and lifts the slices to an SDF with a coordinate-based transformer, to reveal occluded 3D structure.
Problem
Single-view 3D reconstruction struggles to recover occluded parts; multi-view synthesis-based methods suffer from multi-view inconsistency and cannot reveal occluded structure reliably.
Approach
Replace multi-view synthesis with multi-slice prediction: slice the object along its bounding box axes (or camera axes) so that slices reveal occluded parts without camera changes, then reconstruct an implicit SDF from the slice images via a transformer that aggregates per-query features across slices.
Key results
- On ShapeNet Chairs, regression-based Slice3D achieves CD 2.80 and F1 2.43 vs. DISN's 3.25 and 2.34 (ground-truth-free poses).
- On Objaverse, Slice3D (regression) reaches CD 16.6 and LPIPS 12.8 vs. One-2-3-45's CD 35.4 and LPIPS 19.8.
- Inference takes 10-20 seconds on a single Nvidia A40 GPU, versus 90 minutes for RealFusion and 45 seconds for One-2-3-45.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use regression-based rather than generation-based slicing (ablation) | Slice3D · Objaverse | CD: 25.0 -> 16.6 | Helped |
| use 8 slices per direction instead of 4 (ablation) | Slice3D · ShapeNet Chairs | CD: 2.67 -> 2.56 | Helped |
| use one-hot instead of learned slice indicator codes (ablation) | Slice3D · ShapeNet Chairs | CD: 2.67 -> 2.68 | Hurt |
| add estimated depth images to DISN input (ablation) | DISN · ShapeNet Chairs | CD: 3.25 -> 2.99 | Helped |
| use slices along X, Y, and Z instead of only X (ablation) | Slice3D · ShapeNet Chairs | CD: 2.88 -> 2.67 | Helped |
| keep holes in slices instead of filling them (ablation) | Slice3D · ShapeNet Chairs | CD: 2.73 -> 2.67 | Helped |
| remove textures from slice images (ablation) | Slice3D · ShapeNet Chairs | CD: 2.82 -> 2.67 | Helped |
Surprising findings
- Keeping holes (hollow slices) in slice images works better than filling them, since back faces reveal more inner/outer surface information.
- Slice images without textures give better reconstruction than textured slices.
Limitations
- Only four slices per direction is coarse, making far-away parts small and hard to feature-extract.
- Slicing can only be executed on digital 3D models, not physical objects.
- Generation-based slicing requires much longer inference time than regression.
Experimental setup
Datasets: ShapeNet Objaverse GSO
Benchmarks: ShapeNet Chairs ShapeNet Airplanes Objaverse GSO
Metrics: CD F1 HD PSNR SSIM LPIPS
Baselines: DISN AutoSDF NeRF-From-Image SSDNeRF RealFusion One-2-3-45
Builds on: DISN DDPM Zero-1-to-3 One-2-3-45 AutoSDF
Closest papers by meaning
- Sparse3D: Distilling Multiview-Consistent Diffusion for Object Reconstruction from Sparse Views (AAAI 2024 · similarity 0.87)
- Sharp-It: A Multi-view to Multi-view Diffusion Model for 3D Synthesis and Manipulation (CVPR 2025 · similarity 0.87)
- Object-X: Learning to Reconstruct Multi-Modal 3D Object Representations (NeurIPS 2025 · similarity 0.87)
- Novel View Synthesis from A Few Glimpses via Test-Time Natural Video Completion (NeurIPS 2025 · similarity 0.87)
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.86)
- Coherent 3D Scene Diffusion From a Single RGB Image (NeurIPS 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.