Research map / 3D vision / Gaussian splatting
Seeing A 3D World in A Grain of Sand
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 single-camera planar catadioptric lens with eight mirror pairs for snapshot full-surround capture of miniature scenes, and reconstructs them with 3DGS regularized by a visual-hull-derived weighted depth loss.
Problem
3D reconstruction and novel view synthesis of miniature scenes with millimeter-scale objects, which are hard to capture due to shallow depth of field, little texture, and failure of SfM/COLMAP calibration; sparse surround views also degrade 3DGS.
Approach
A one-bounce planar catadioptric mirror-pair lens folds the light path to act as a zoom lens and yields eight time-synchronized virtual views with pre-calibrated poses; a weighted depth loss from visual hulls (penalizing depths beyond the hull with an S-shaped weight) regularizes sparse-view 3DGS.
Key results
- On synthetic data, the proposed method reaches PSNR 32.48, SSIM 0.9783, LPIPS 0.0265, outperforming SparseGS (31.84 PSNR) and Hierarchical 3DGS (26.83 PSNR).
- FSGS scores PSNR 18.93 and DNGaussian 21.40 on the same synthetic benchmark, both well below the proposed method.
- Reconstruction of a scene from 8 reference views at 800x800 takes about 2 minutes on an NVIDIA 4090 GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add visual hull-constrained depth regularization to sparse-view 3DGS | 3DGS miniature-scene reconstruction · Synthetic scenes (24 novel views per scene) | SSIM: 0.9756 -> 0.9783 | Helped |
| Add visual hull-constrained depth regularization to sparse-view 3DGS | 3DGS miniature-scene reconstruction · Synthetic scenes (24 novel views per scene) | PSNR: 31.8415 -> 32.4792 | Helped |
| Add visual hull-constrained depth regularization to sparse-view 3DGS | 3DGS miniature-scene reconstruction · Synthetic scenes (24 novel views per scene) | LPIPS: 0.0298 -> 0.0265 | Helped |
Surprising findings
- COLMAP fails to run on the miniature scenes due to little texture, so pre-calibrated camera parameters are required for all compared methods.
Limitations
- Angular resolution is reduced compared with kaleidoscopic systems because only one-bounce reflections are used, giving fewer viewing directions.
- Evaluated on a small number of synthetic scenes (7) and a limited set of real scenes (12), with real-scene evaluation mostly visual.
Experimental setup
Datasets: synthetic miniature scenes rendered in Autodesk 3ds Max (7 scenes) real miniature scenes captured with custom mirror lens (10 indoor, 2 outdoor)
Benchmarks: synthetic miniature scenes (7 scenes, 24 novel views each) real captured miniature scenes
Metrics: SSIM PSNR LPIPS
Baselines: Hierarchical 3DGS FSGS DNGaussian SparseGS InstantSplat
Builds on: 3D Gaussian Splatting Hierarchical 3DGS Folded optics Segment Anything
Closest papers by meaning
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.86)
- CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction (NeurIPS 2026 · similarity 0.86)
- HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction (ICLR 2025 · similarity 0.85)
- GigaGS: 3D Gaussian Based Planar Representation for Large-Scene Surface Reconstruction (AAAI 2025 · similarity 0.85)
- No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images (ICLR 2025 · similarity 0.85)
- 3D Gaussian Flats: Hybrid 2D/3D Photometric Scene Reconstruction (NeurIPS 2025 · 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.