Research map / 3D vision / Depth and stereo
High Dynamic Range Novel View Synthesis with Single Exposure
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 single-exposure HDR novel view synthesis and Mono-HDR-3D, a meta-framework with LDR-to-HDR and closed-loop HDR-to-LDR color converters that plugs into NeRF and 3DGS to learn HDR scenes from single-exposure LDR images.
Problem
Learning an HDR 3D scene model for novel view synthesis from LDR images captured at a single exposure, avoiding the motion artifacts and capture costs of multi-exposure HDR-NVS.
Approach
Model the camera imaging formation (exposure scaling, offset, saturation, noise) with two learned converters: L2H-CC lifts LDR renderings to HDR, and H2L-CC maps HDR renderings back to LDR so the model can be supervised by LDR images in a closed loop, usable with NeRF or 3DGS backbones.
Key results
- On synthetic data, Mono-HDR-GS reaches 38.57 dB HDR PSNR versus 35.30 dB for HDR-GS (SSIM 0.975 vs 0.965).
- Mono-HDR-NeRF reaches 32.86 dB HDR PSNR versus 13.76 dB for HDR-NeRF under the single-exposure setting.
- On real scenes, Mono-HDR-GS improves LDR PSNR to 35.81 dB from 35.34 dB for HDR-GS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Mono-HDR-3D to NeRF | HDR-NeRF / NeRF · Synthetic datasets | PSNR (HDR): 13.76 -> 32.86 | Helped |
| Add Mono-HDR-3D to 3DGS | HDR-GS / 3DGS · Synthetic datasets | PSNR (HDR): 35.30 -> 38.57 | Helped |
| Add Mono-HDR-3D to NeRF | HDR-NeRF / NeRF · Real datasets | PSNR (LDR): 32.50 -> 32.52 | No clear effect |
| Add Mono-HDR-3D to 3DGS | HDR-GS / 3DGS · Real datasets | PSNR (LDR): 35.34 -> 35.81 | Helped |
| Use camera-imaging-aware L2H-CC rather than an MLP (ablation) | Mono-HDR-GS · Synthetic datasets | PSNR (HDR): 19.02 -> 38.57 | Helped |
| Use camera-imaging-aware H2L-CC rather than an MLP (ablation) | Mono-HDR-GS · Synthetic datasets | PSNR (HDR): 38.43 -> 38.57 | Helped |
| Add H2L-CC closed-loop design (ablation) | Mono-HDR-GS · Synthetic datasets | PSNR (HDR): 38.19 -> 38.57; 0.38 dB | Helped |
| Use L_hdr loss rather than L_h2l alone (ablation) | Mono-HDR-GS · Synthetic datasets | PSNR (HDR): 11.87 -> 33.93 | Helped |
| Add L_hdr loss to L_ldr + L_h2l (ablation) | Mono-HDR-GS · Synthetic datasets | PSNR (HDR): 13.50 -> 38.57 | Helped |
| Add L_ldr loss to L_hdr (ablation) | Mono-HDR-GS · Synthetic datasets | PSNR (HDR): +4.26 dB; 33.93 -> 38.19 | Helped |
| Add L_h2l loss to L_ldr + L_hdr (ablation) | Mono-HDR-GS · Synthetic datasets | PSNR (HDR): +0.38 dB; 38.19 -> 38.57 | Helped |
| Use L_ldr loss alone (ablation) | Mono-HDR-GS · Synthetic datasets | PSNR (HDR) | Hurt |
Surprising findings
- HDR-NeRF fails to converge with single-exposure data, often producing entirely black or white images.
- Mono-HDR-GS trained only on LDR images (1/0 ratio) slightly outperforms HDR-GS trained solely on HDR images.
Limitations
- Evaluated only on the small HDR-NeRF synthetic and real datasets with a fixed set of exposure times.
- Real scenes lack ground-truth HDR, so HDR quality cannot be measured on real data.
- Performance on real-world diverse environments is left to future work.
Experimental setup
Datasets: HDR-NeRF synthetic multi-exposure dataset (Blender, 8 scenes) HDR-NeRF real dataset (4 scenes)
Benchmarks: HDR-NeRF synthetic scenes HDR-NeRF real scenes
Metrics: PSNR SSIM LPIPS fps
Baselines: HDR-NeRF HDR-GS
Builds on: HDR-NeRF HDR-GS NeRF 3D Gaussian Splatting
Closest papers by meaning
- HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting (NeurIPS 2024 · similarity 0.92)
- EvHDR-NeRF: Building High Dynamic Range Radiance Fields with Single Exposure Images and Events (AAAI 2025 · similarity 0.89)
- Towards Real-World HDR Video Reconstruction: A Large-Scale Benchmark Dataset and A Two-Stage Alignment Network (CVPR 2024 · similarity 0.89)
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.89)
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.88)
- Perceptual Assessment and Optimization of HDR Image Rendering (CVPR 2024 · similarity 0.88)
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.