Research map / Video / Video segmentation and tracking
HDRFlow: Real-Time HDR Video Reconstruction with Large Motions
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 HDRFlow, a fast HDR-oriented optical flow estimator with multi-size large-kernel convolutions, an HDR-domain alignment loss, and synthetic Sintel training data for real-time HDR video reconstruction from alternating exposures.
Problem
HDR video reconstruction from alternating-exposure LDR frames suffers from ghosting under large camera or object motion, and existing flow- or attention-based alignment is slow and fails on large motions and saturated or dark regions.
Approach
A lightweight single-pass flow network with depthwise multi-size large kernels (7x7, 9x9, 11x11) at the coarsest scale, supervised by an HDR-domain photometric loss computed on tonemapped clean HDR frames and masked to under/over-exposed regions, plus Sintel forward/backward flow supervision.
Key results
- On Cinematic Video, Ours (Vimeo+Sintel) reaches 39.30 PSNR-T for 2-exposure and 36.65 for 3-exposure, up to 1.08 dB and 1.58 dB above the second-best method.
- Flow network plus fusion runs in 55 ms (2-exposure) on 1920x1080 Cinematic Video, roughly 10x faster than Chen and LAN-HDR.
- Compared with RAFT+fusion, HDRFlow gets 45.50 PSNR-T on DeepHDRVideo-D versus 44.85 and runs about 15x faster (35 ms vs 532 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the proposed flow network instead of SPyNet (ablation) | HDR video reconstruction · DeepHDRVideo dynamic scenes | PSNRT: 44.82 -> 45.10 | Helped |
| Add multi-size large kernel (MLK) (ablation) | HDRFlow flow network · DeepHDRVideo dynamic scenes | PSNRT: 45.10 -> 45.30 | Helped |
| Add HDR-domain alignment loss (HALoss) (ablation) | HDRFlow · DeepHDRVideo dynamic scenes | PSNRT: 45.30 -> 45.41 | Helped |
| Add luminance mask to restrict HALoss to poorly exposed regions (ablation) | HDRFlow · DeepHDRVideo dynamic scenes | PSNRT: 45.41 -> 45.50 | Helped |
| Train with Sintel data and flow supervision in addition to Vimeo (ablation) | HDRFlow · Cinematic Video, 2-Exposure | PSNRT: 39.20 -> 39.30 | Helped |
| Train with Sintel data and flow supervision in addition to Vimeo (ablation) | HDRFlow · Cinematic Video, 3-Exposure | PSNRT: 36.55 -> 36.65 | Helped |
| Train with Sintel data and flow supervision in addition to Vimeo (ablation) | HDRFlow · DeepHDRVideo, 2-Exposure | PSNRT: 43.18 -> 43.25 | Helped |
| Train with Sintel data and flow supervision in addition to Vimeo (ablation) | HDRFlow · DeepHDRVideo, 3-Exposure | PSNRT: 40.45 -> 40.56 | Helped |
| Replace the pretrained RAFT flow estimator with HDRFlow | HDR video reconstruction · DeepHDRVideo dynamic scenes | PSNRT: 44.85 -> 45.50 | Helped |
Surprising findings
- Sharp flow boundaries are not necessary for HDR fusion; the fused output stays sharp even though the flow is less crisp than RAFT's.
Limitations
- Flow boundaries are less sharp than those from expensive iterative or transformer flow methods.
- Evaluation relies on synthetically converted LDR sequences rather than real captured alternating-exposure video for training.
Experimental setup
Datasets: Vimeo-90K Sintel Cinematic Video DeepHDRVideo Kalantari13 HDRVideo
Benchmarks: Cinematic Video DeepHDRVideo
Metrics: PSNR-T SSIM-T HDR-VDP-2 inference time (ms)
Baselines: Kalantari13 Kalantari19 Yan19 Prabhakar Chen (DeepHDRVideo) LAN-HDR RAFT+fusion Base (SPyNet)
Builds on: SpyNet RAFT Kalantari19 HDR video network DeepHDRVideo LAN-HDR
Closest papers by meaning
- Towards Real-World HDR Video Reconstruction: A Large-Scale Benchmark Dataset and A Two-Stage Alignment Network (CVPR 2024 · similarity 0.92)
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.88)
- EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting (AAAI 2025 · 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)
- Generating Content for HDR Deghosting from Frequency View (CVPR 2024 · similarity 0.86)
- HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting (NeurIPS 2024 · similarity 0.86)
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.