Research map / Image generation / Image restoration
Dynamic Exposure Burst Image Restoration
Woohyeok Kim, Jaesung Rim, Daeyeon Kim, Sunghyun Cho
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| predict per-frame exposure times with BAENet instead of fixed bracket | DEBIR/Burstormer · Simulated test set | PSNR: 35.04 -> 35.32 | Improved |
| predict per-frame exposure times instead of real-camera exposure bracketing | DEBIR · Real-world captures | BRISQUE: 50.66 -> 46.90 | Improved |
| use non-uniform instead of uniform exposure times (ablation) | DEBIR · Simulated test set | PSNR: 35.01 -> 35.32 | Improved |
| add preview image input to BAENet (ablation) | BAENet · Simulated test set | PSNR: 34.80 -> 35.32 | Improved |
| add gain input to BAENet (ablation) | BAENet · Simulated test set | PSNR: 35.21 -> 35.32 | Improved |
| add motion magnitude input to BAENet (ablation) | BAENet · Simulated test set | PSNR: 35.13 -> 35.32 | Improved |
| add BAENet warm-up with pseudo-ground-truth exposure times (S2-1) (ablation) | DEBIR training · Simulated test set | PSNR: 35.16 -> 35.32 | Improved |
| add fine-tuning of restoration network (S3) (ablation) | DEBIR training · Simulated test set | PSNR: 35.01 -> 35.32 | Improved |
| use staged training instead of end-to-end training (ablation) | DEBIR training · Simulated test set | PSNR: 33.60 -> 35.32 | Improved |
| use end-to-end finetuning instead of S3 restoration finetuning (ablation) | DEBIR training · Simulated test set | PSNR: 35.32 -> 35.11 | Hurt |
| increase burst size from 2 to 8 (ablation) | DEBIR · Simulated test set | PSNR: 34.22 -> 36.11 | Improved |
| DEBIR vs exposure bracketing as burst size grows (ablation) | DEBIR · Simulated test set | PSNR: +0.63 dB at n=2, +0.22 at n=8 (34.22 vs 33.59; 36.11 vs 35.89) | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Digital-Gimbal — Digital Gimbal: End-to-End Deep Image Stabilization With Learnable Exposure Times (CVPR 2021) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Burst Image Restoration and Enhancement (CVPR 2022 · similarity 0.88)
- Burstormer: Burst Image Restoration and Enhancement Transformer (CVPR 2023 · similarity 0.87)
- Zero-Shot Burst Restoration via Diffusion MAP Inference with Poisson-Gaussian Noise Likelihood (NeurIPS 2026 · similarity 0.86)
- Iterative Residual CNNs for Burst Photography Applications (CVPR 2019 · similarity 0.86)
- Burst Image Deblurring Using Permutation Invariant Convolutional Neural Networks (ECCV 2018 · similarity 0.86)
- Exposure Bracketing Is All You Need For A High-Quality Image (ICLR 2025 · similarity 0.83)
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.
