Research map / 3D vision / Gaussian splatting
EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting
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.
Builds a self-supervised, event-guided HDR video reconstruction method by fitting HDR 3D Gaussians to single-exposure LDR frames and events, with a learnable HDR-to-LDR CRF transform.
Problem
HDR video reconstruction from single-exposure LDR video suffers from brightness inconsistency across frames, and supervised learning methods suffer from data bias that causes color shifts (e.g., yellow tint) on out-of-distribution inputs.
Approach
Treat the LDR video frames as multi-views of a static scene and optimize HDR 3D Gaussians (each with an HDR and an LDR color branch) so 3D consistency enforces temporal brightness consistency; a learnable exponential CRF maps HDR to LDR and is supervised by event streams and LDR frames, with no pre-collected training data.
Key results
- On synthetic data, EvHDR-GS reaches PSNR 27.94, SSIM 0.971, LPIPS 0.037, and VDP 8.651, the best on all metrics versus Cui et al. (25.13 PSNR) and Chung and Cho (25.42 PSNR).
- Removing the event loss drops PSNR to 17.14, showing events are essential to recover HDR information.
- Optimization takes about 7 minutes per 100-frame video, while rendering takes 10.4 ms per frame on a single A5000.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EvHDR-GS instead of Chung and Cho | HDR video reconstruction · Synthetic data | PSNR: 25.42 -> 27.94 | Helped |
| use EvHDR-GS instead of Yang et al. | HDR video reconstruction · Synthetic data | SSIM: 0.873 -> 0.971 | Helped |
| use EvHDR-GS instead of Chung and Cho | HDR video reconstruction · Synthetic data | LPIPS: 0.089 -> 0.037 | Helped |
| use EvHDR-GS instead of Chung and Cho | HDR video reconstruction · Synthetic data | VDP: 8.356 -> 8.651 | Helped |
| add event loss (ablation) | EvHDR-GS · Synthetic data | PSNR: 17.14 -> 27.94 | Helped |
| add event loss (ablation) | EvHDR-GS · Synthetic data | SSIM: 0.680 -> 0.971 | Helped |
| add event loss (ablation) | EvHDR-GS · Synthetic data | LPIPS: 0.257 -> 0.037 | Helped |
| add event loss (ablation) | EvHDR-GS · Synthetic data | VDP: 6.136 -> 8.651 | Helped |
Surprising findings
- Without event supervision, the 3DGS-based variant performs worse than most existing HDR reconstruction baselines in PSNR (17.14 vs. 18.44 to 25.42).
Limitations
- Limited to static scenes because vanilla 3DGS represents only static geometry.
- Requires per-scene optimization, which is slower than feed-forward baselines at setup time.
- Relies on SfM/COLMAP camera poses for every frame.
Experimental setup
Datasets: DeepHDRVideo-Dataset HDR-NeRF Blender scenes HDR-plenoxels Blender scenes self-collected real-world DAVIS 346C event and image videos
Benchmarks: synthetic HDR video scenes (DeepHDRVideo, HDR-NeRF, HDR-plenoxels Blender) real-world DAVIS 346C captures
Metrics: PSNR SSIM LPIPS HDR-VDP-3 inference time
Baselines: Liu et al. 2020 Chung and Cho 2023 Xu et al. 2024 Cui et al. 2024a Rebecq et al. 2019b Yang et al. 2023
Builds on: 3D Gaussian Splatting event generation model (v2e-style log-brightness model) Yang et al. 2023 event-guided HDR video COLMAP / SfM
Closest papers by meaning
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.91)
- HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- From Chaos to Clarity: 3DGS in the Dark (NeurIPS 2024 · similarity 0.90)
- Towards Real-World HDR Video Reconstruction: A Large-Scale Benchmark Dataset and A Two-Stage Alignment Network (CVPR 2024 · similarity 0.89)
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.88)
- EvHDR-NeRF: Building High Dynamic Range Radiance Fields with Single Exposure Images and Events (AAAI 2025 · 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.