Research map / Video / Video segmentation and tracking
Towards Real-World HDR Video Reconstruction: A Large-Scale Benchmark Dataset and A Two-Stage Alignment Network
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 Real-HDRV, a 500-video real-world HDR video dataset with per-frame HDR labels, and a two-stage (global offset + local adaptive separable convolution) alignment network that outperforms prior HDR video methods.
Problem
HDR video reconstruction from alternating-exposure LDR sequences is limited by the lack of large-scale real-world training data (existing models train on synthetic data that generalizes poorly) and by alignment errors under global and local motion with noise and saturation.
Approach
Captures real LDR-HDR video pairs frame by frame using 7-exposure stacks at high burst speed to get per-frame HDR labels, and aligns inputs sequentially: a global module estimating weighted pre-defined offset bases (8 DoF) followed by a coarse-to-fine local module using adaptive separable convolution guided by exposure masks.
Key results
- Proposed method achieves average PSNR-µ 43.86 dB on Real-HDRV, the best among compared methods (Yue23 43.48, Chen21 42.12).
- Models trained on Real-HDRV outperform those trained on the synthetic dataset; e.g., Kalantari19 gains more than 1 dB PSNR-µ on the Chen21 dynamic set.
- Ablation: removing GAM and LAM lowers PSNR-µ from 43.86 to 42.67 (baseline) and HDR-VDP-2 from 63.53 to 59.49.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train Kalantari19 on Real-HDRV instead of the synthetic dataset | Kalantari19 · Chen21 dynamic set | PSNR-µ: 44.15 -> 45.31 | Helped |
| Train Kalantari19 on Real-HDRV instead of the synthetic dataset | Kalantari19 · Chen21 static set | PSNR-µ: 40.59 -> 41.19 | Helped |
| Add the global alignment module (GAM) (ablation) | HDR video reconstruction network · Real-HDRV | HDR-VDP-2: 59.49 -> 61.72 | Helped |
| Add the global alignment module (GAM) (ablation) | HDR video reconstruction network · Real-HDRV | PSNR-µ: 42.67 -> 43.07 | Helped |
| Add the global alignment module (GAM) (ablation) | HDR video reconstruction network · Real-HDRV | HDR-VQM: 85.34 -> 87.62 | Helped |
| Add the local alignment module (LAM) (ablation) | HDR video reconstruction network · Real-HDRV | HDR-VDP-2: 59.49 -> 62.58 | Helped |
| Add the local alignment module (LAM) (ablation) | HDR video reconstruction network · Real-HDRV | PSNR-µ: 42.67 -> 43.61 | Helped |
| Add the local alignment module (LAM) (ablation) | HDR video reconstruction network · Real-HDRV | HDR-VQM: 85.34 -> 88.45 | Helped |
| Add GAM to the network with LAM (ablation) | HDR video reconstruction network · Real-HDRV | HDR-VDP-2: 62.58 -> 63.53 | Helped |
| Add GAM to the network with LAM (ablation) | HDR video reconstruction network · Real-HDRV | PSNR-µ: 43.61 -> 43.86 | Helped |
| Add GAM to the network with LAM (ablation) | HDR video reconstruction network · Real-HDRV | HDR-VQM: 88.45 -> 89.71 | Helped |
Surprising findings
- Trained on the synthetic dataset, several existing methods show color distortions and ghosting on real scenes, whereas training on Real-HDRV transfers well to Chen21.
Limitations
- Test set is relatively small (50 videos), and dataset scenes are carefully selected static scenes with manually created motion.
- Ground truth for dynamic content relies on per-frame multi-exposure stacks, so the dataset is limited to quasi-static captures with manual motion.
Experimental setup
Datasets: Real-HDRV Chen21 dataset synthetic dataset from Chen21 (21 HDR videos from Kalantari/Froehlich-style sources and Vimeo-90K)
Benchmarks: Real-HDRV test set Chen21 dataset
Metrics: PSNR-µ SSIM-µ PU-PSNR PU-SSIM HDR-VDP-2 HDR-VQM
Baselines: Kalantari13 AHDRNet Kalantari19 Chen21 CA-ViT Yue23 LAN-HDR
Builds on: AHDRNet Kalantari19 HDR video LAN-HDR Mertens exposure fusion
Closest papers by meaning
- HDRFlow: Real-Time HDR Video Reconstruction with Large Motions (CVPR 2024 · similarity 0.92)
- EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting (AAAI 2025 · similarity 0.89)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.89)
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.88)
- HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting (NeurIPS 2024 · similarity 0.88)
- High Dynamic Range Video Compression: A Large-Scale Benchmark Dataset and A Learned Bit-depth Scalable Compression Algorithm (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.