Research map / 3D vision / NeRF and radiance fields
ESR-NeRF: Emissive Source Reconstruction Using LDR Multi-view Images
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.
Introduces ESR-NeRF, a NeRF-based inverse rendering method that reconstructs any number of emissive sources from LDR multi-view images captured with lights on and off, using learned light-transport-segment losses and reflection-aware progressive refinement.
Problem
NeRF-based inverse rendering assumes distant lighting and ignores emissive sources in a scene; recovering emissive sources from LDR images is ambiguous due to limited dynamic range and expensive multi-bounce volume rendering.
Approach
Uses neural networks as learnable functions for ray-traced fields, enforcing light transport segment (LTS) consistency on separate outgoing radiance fields for emission-off and emission-on components, and progressively separates emission from reflection by splitting rays into uncertain and certain groups, with a learnable tone-mapper to resolve HDR/LDR ambiguity.
Key results
- ESR-NeRF achieves IoU 0.81 on white-colored Lego scene versus 0.71 for TensoIR and 0.43 for NeILF++ in emissive source identification.
- Scene editing PSNR with intensity and color editing reaches 31.63 on Twins-style setting versus 22.58 for PaletteNeRF; ESR-NeRF best on all editing settings.
- On DTU surface reconstruction, ESR-NeRF achieves mean Chamfer distance 0.65, versus 0.68 for Neural-PBIR and 0.77 for NeuS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add progressive refinement of emissive sources with reflection awareness (ablation) | ESR-NeRF · White-colored source identification | IoU: 0.40 -> 0.86 | Helped |
| Add progressive refinement of emissive sources with reflection awareness (ablation) | ESR-NeRF · White-colored source intensity restoration | MSE: 9.92 -> 5.24 | Helped |
| Add stop-gradient operation (ablation) | ESR-NeRF · White source identification | IoU: 0.71 -> 0.86 | Helped |
| Add stop-gradient operation (ablation) | ESR-NeRF · White source intensity restoration | MSE: 6.45 -> 5.24 | Helped |
| Add learnable tone-mapper (ablation) | ESR-NeRF · DTU | Chamfer distance: 0.93 -> 0.65 | Helped |
| Add light transport segment loss (ablation) | ESR-NeRF · DTU | Chamfer distance: 0.71 -> 0.65 | Helped |
Surprising findings
- Standard NeRF-based relighting methods such as TensoIR and NeILF++ achieve near-zero IoU on emissive source identification due to source ambiguity.
- Learnable tone-mapper removal degrades surface reconstruction on DTU (CD 0.93 vs 0.65).
Limitations
- LTS-based relighting may be weak in representing new colors that traverse unobserved light paths during training.
- Requires emissive sources toggled on and off; volume ray tracing in unbounded scenes not addressed.
- Struggles to capture complex reflections within light bulbs in real scenes.
Experimental setup
Datasets: DTU
Benchmarks: DTU custom emissive source synthetic scenes
Metrics: IoU MSE PSNR LPIPS Chamfer distance
Baselines: TensoIR NeILF++ Twins NeRF-W PaletteNeRF NeuS Voxurf Neural-PBIR
Builds on: NeRF Voxurf NeuS Disney BRDF
Closest papers by meaning
- Pano-NeRF: Synthesizing High Dynamic Range Novel Views with Geometry from Sparse Low Dynamic Range Panoramic Images (AAAI 2024 · similarity 0.89)
- Inverse Rendering of Glossy Objects via the Neural Plenoptic Function and Radiance Fields (CVPR 2024 · similarity 0.89)
- Aleth-NeRF: Illumination Adaptive NeRF with Concealing Field Assumption (AAAI 2024 · similarity 0.89)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.88)
- Optimize the Unseen - Fast NeRF Cleanup with Free Space Prior (NeurIPS 2025 · similarity 0.87)
- PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields (CVPR 2025 · similarity 0.87)
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.