Research map / 3D vision / Depth and stereo
OmniSDF: Scene Reconstruction using Omnidirectional Signed Distance Functions and Adaptive Binoctrees
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.
Reconstructs static indoor/outdoor scene geometry from short circular-sweep omnidirectional video by optimizing a NeuS-style SDF guided by an adaptive spherical binoctree (sphoxels) that is subdivided during training.
Problem
Neural SDF reconstruction from omnidirectional video captured with a small circular camera sweep is hard due to small baseline, large depth range, and wide per-pixel angular extent; fixed-resolution grids are memory-inefficient and undersample detail.
Approach
Use a spherical binoctree (sphoxels) to guide ray sampling for a NeuS SDF, with coarse sphoxels dilated from initial depth and fine sphoxels adaptively subdivided during optimization near the surface, plus a BFS-based traversal and a sphoxel intersection test.
Key results
- Average inverse-depth RMSE of 0.69 across the three synthetic scenes, versus 0.84 for COLMAP and 0.88 for EgocentricRecon (Table 2).
- Best neural baseline Neuralangelo averages RMSE 1.16 and NeuS-facto 2.91; the method improves on EgocentricRecon's binoctree approach by about 38% on average per the introduction.
- Sponza requires about 4.3M sphoxels versus 33.3B voxels for a dense regular grid at matching minimum size (Table 1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive spherical binoctree sampling instead of a dense Cartesian grid | OmniSDF · Sponza, Lone-monk, San Miguel | Number of voxels: 33,335,054,331 vs 4,346,041; 2,234,638,740 vs 231,237; 1,953,273,076 vs 951,703 | Helped |
| Use OmniSDF instead of COLMAP | OmniSDF · Average | RMSE: 0.84 -> 0.69 | Helped |
| Use OmniSDF instead of COLMAP | OmniSDF · Lone-monk | RMSE: 0.40 -> 0.69 | Hurt |
| Use OmniSDF instead of Neuralangelo | OmniSDF · San Miguel | RMSE: 0.58 -> 0.56 | Helped |
| Use OmniSDF instead of COLMAP | OmniSDF · Average | MAE: 0.28 -> 0.32 | Hurt |
| Use binoctree guidance instead of adding direct depth supervision | OmniSDF vs NeuS-facto-D · Average | RMSE: 2.67 -> 0.69 | Helped |
| Add coarse sphoxel sampling to sphere sampling (ablation) | OmniSDF · Sponza | reconstruction quality | Helped |
| Add adaptive fine sphoxel sampling to sphere and coarse sampling (ablation) | OmniSDF · Sponza | reconstruction quality | Helped |
Surprising findings
- Depth supervision (NeuS-facto-D) gave only modest gains over plain NeuS-facto (average RMSE 2.67 vs 2.91), and the binoctree-guided sampling was better than direct depth supervision.
Limitations
- Relies on initial depth estimates for coarse voxels; wrong depth can misguide geometry and background regions with near initial depth.
- Concave regions are sensitive to optimization minima.
- Fine-scale texture variation may be baked into geometry.
Experimental setup
Datasets: Blender Online Community scenes (Sponza, Lone-monk, San Miguel) McGuire Computer Graphics Archive scenes Omniphotos dataset
Benchmarks: Sponza Lone-monk San Miguel
Metrics: RMSE MAE voxel count
Baselines: COLMAP EgocentricRecon NeuS NeuS-facto Neuralangelo NeuS-facto-D
Builds on: NeuS Egocentric binoctree reconstruction (Jang et al.) mip-NeRF 360 proposal sampling Instant-NGP
Closest papers by meaning
- OmniLocalRF: Omnidirectional Local Radiance Fields from Dynamic Videos (CVPR 2024 · similarity 0.86)
- ODGS: 3D Scene Reconstruction from Omnidirectional Images with 3D Gaussian Splattings (NeurIPS 2024 · similarity 0.86)
- SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting (ICLR 2025 · similarity 0.85)
- CRAYM: Neural Field Optimization via Camera RAY Matching (NeurIPS 2024 · similarity 0.85)
- Omni-Scene: Omni-Gaussian Representation for Ego-Centric Sparse-View Scene Reconstruction (CVPR 2025 · similarity 0.84)
- OmniSplat: Taming Feed-Forward 3D Gaussian Splatting for Omnidirectional Images with Editable Capabilities (CVPR 2025 · 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.