Research map / 3D vision / Depth and stereo
LightOctree: Lightweight 3D Spatially-Coherent Indoor Lighting Estimation
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 LightOctree, a lightweight voxel-octree-based network with a differentiable cone-tracing rendering layer that estimates spatially-coherent 3D indoor lighting from a single RGB image at much lower memory and compute cost.
Problem
Estimating spatially-coherent 3D indoor lighting from a single narrow-FOV RGB image with volumetric representations is memory- and compute-intensive because it ignores the sparsity of light sources in space.
Approach
Represent the lighting as a sparse voxel octree (RGBα per non-empty node) predicted by an octree U-Net (graph O-CNN operators), and render supervision with a differentiable multi-scale cone-tracing layer that samples octree depths according to distance, avoiding dense 3D grids.
Key results
- On InteriorNet, the method reaches 16.82 dB PSNR, above Gardner et al. (13.42) and NIR (16.59) but below Lighthouse (17.29) and Wang et al. (17.37).
- On FutureHouse, the spatial-coherence SC metric is 0.226, better than Gardner (0.291), Li et al. (0.274), and Lighthouse (0.243).
- In a 20-participant user study, the method was preferred over Gardner, Li et al., and Lighthouse (65–75% response, p<0.001).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use octree depth 7 instead of depth 6 (ablation) | LightOctree · FutureHouse | PSNR: 15.34 -> 17.14 | Helped |
| use GT depth map with octree depth 6 (ablation) | LightOctree · FutureHouse | PSNR: 15.34 -> 15.56 | Helped |
| use GT depth map with octree depth 7 (ablation) | LightOctree · FutureHouse | PSNR: 17.14 -> 17.21 | Helped |
| use Spherical Gaussian lighting representation with octree depth 6 (ablation) | LightOctree · FutureHouse | PSNR: 15.34 -> 15.62 | Helped |
| estimate lighting with LightOctree rather than Lighthouse | indoor lighting estimation · InteriorNet | PSNR (dB): 17.29 -> 16.82 | Hurt |
| estimate lighting with LightOctree rather than Wang et al. | indoor lighting estimation · InteriorNet | PSNR (dB): 17.37 -> 16.82 | Hurt |
| use LightOctree instead of Gardner et al. | spatially-coherent lighting estimation · FutureHouse | SC metric: 0.291 -> 0.226 | Helped |
| use LightOctree instead of Lighthouse | spatially-coherent lighting estimation · FutureHouse | SC metric: 0.243 -> 0.226 | Helped |
Surprising findings
- Despite its efficiency focus, the method trails SOTA in PSNR, which the authors acknowledge.
Limitations
- Lower PSNR than state-of-the-art methods such as Lighthouse and Wang et al.
- Trained only on synthetic FutureHouse data; relies on depth prediction inputs.
Experimental setup
Datasets: FutureHouse synthetic indoor dataset InteriorNet Laval Indoor
Benchmarks: InteriorNet FutureHouse Laval Indoor
Metrics: PSNR SC loss (spatial coherence) user study preference inference time VRAM usage
Baselines: Gardner et al. NIR DeepLight Garon et al. Li et al. Lighthouse Wang et al.
Builds on: Lighthouse O-CNN PhyIR DenseNet
Closest papers by meaning
- Fast Spatially-Varying Indoor Lighting Estimation (CVPR 2019 · similarity 0.90)
- Neural Illumination: Lighting Prediction for Indoor Environments (CVPR 2019 · similarity 0.87)
- Lighthouse: Predicting Lighting Volumes for Spatially-Coherent Illumination (CVPR 2020 · similarity 0.87)
- ScribbleLight: Single Image Indoor Relighting with Scribbles (CVPR 2025 · similarity 0.85)
- Sparse Views Near Light: A Practical Paradigm for Uncalibrated Point-light Photometric Stereo (CVPR 2024 · similarity 0.83)
- SpecNeRF: Gaussian Directional Encoding for Specular Reflections (CVPR 2024 · similarity 0.82)
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.