Research map / Detection and segmentation / Open-set and few-shot detection
ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge
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 ReRAW, a multi-head gamma-space RGB-to-RAW reverse ISP trained with stratified patch sampling and a hard-log loss, and shows that pretraining small object detectors on ReRAW-generated synthetic RAW plus real RAW beats RGB pipelines.
Problem
Labeled RAW datasets for training edge object detectors are scarce and costly, so existing labeled RGB datasets need to be converted into sensor-specific RAW images with high fidelity, especially in bright regions of RAW images.
Approach
A reverse ISP that predicts multiple gamma-corrected RAW candidates via a multi-head predictor, re-linearizes and combines them with an image-dependent scaling vector, trained on stratified-sampled patches balanced by brightness and with a hard-log loss.
Key results
- ReRAW-S reaches 41.00 dB PSNR on NOD-Nikon and 38.88 dB on PASCALRAW, outperforming RAW-Diffusion (39.82 and 35.34 dB).
- Pretraining YOLOX-s on BDD-ReRAW-S then fine-tuning on GT RAW raises PASCALRAW mAP from 65.36 (RGB pipeline) to 65.85 and NOD-Nikon mAP from 27.14 to 29.03.
- Hard-log loss gives 30.45 dB on FiveK-Canon versus 29.85 for L1 and 27.93 for L2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use stratified rather than random patch sampling for ReRAW training (ablation) | ReRAW · RGB-to-RAW reconstruction across five datasets | PSNR | Mixed |
| Add the hard-log loss instead of L1 loss (ablation) | ReRAW · RGB-to-RAW reconstruction across five datasets | PSNR | Mixed |
| Add the hard-log loss instead of L2 loss (ablation) | ReRAW · RGB-to-RAW reconstruction across five datasets | PSNR | Helped |
| Add the Global Context Encoder (ablation) | ReRAW · NOD-Nikon and FiveK-Canon | PSNR: 36.04 -> 40.95; 26.55 -> 29.5 | Helped |
| Increase the gamma heads from 1 to 2 without the Gamma Scaling Encoder (ablation) | ReRAW · NOD-Nikon and FiveK-Canon | PSNR: 36.04 -> 36.10; 26.55 -> 27.44 | Helped |
| Add the Gamma Scaling Encoder to the 2-head model (ablation) | ReRAW · NOD-Nikon and FiveK-Canon | PSNR: 36.10 -> 40.87; 27.44 -> 29.75 | Mixed |
| Increase gamma heads from 2 to 10 while using GCE and GSE (ablation) | ReRAW · NOD-Nikon and FiveK-Canon | PSNR: 40.87 -> 41.00; 29.75 -> 30.45 | Mixed |
| Pretrain on ReRAW-S synthetic RAW and fine-tune on GT-RAW instead of the RGB pipeline | RTMDet-s · PASCALRAW | mAP: 57.62 -> 63.19 | Helped |
| Pretrain on ReRAW-S synthetic RAW and fine-tune on GT-RAW instead of the RGB pipeline | RTMDet-s · NOD-Nikon | mAP: 20.46 -> 21.09 | Helped |
| Pretrain on ReRAW-S synthetic RAW and fine-tune on GT-RAW instead of the RGB pipeline | YOLOX-s · PASCALRAW | mAP: 65.36 -> 65.85 | Helped |
| Pretrain on ReRAW-S synthetic RAW and fine-tune on GT-RAW instead of the RGB pipeline | YOLOX-s · NOD-Nikon | mAP: 27.14 -> 29.03 | Helped |
| Pretrain on ReRAW-S synthetic RAW and fine-tune on GT-RAW instead of the RGB pipeline | SSD · PASCALRAW | mAP: 62.50 -> 63.09 | Helped |
Surprising findings
- RTMDet on NOD-Nikon performed better when pretrained on the lower-PSNR random-sampling synthetic set (BDD-ReRAW-R), which the authors attribute to a limitation of PSNR as a proxy for downstream accuracy.
- Fine-tuning RGB-pretrained detectors on RAW (variant b) gives only small or inconsistent gains, and for PASCALRAW the RGB pipeline remains second best.
Limitations
- PSNR does not always predict downstream detection accuracy, as seen with RTMDet on NOD-Nikon.
- Detection experiments use only three small single-stage detectors and two target datasets.
- Large RAW datasets are not available, so the reverse ISP must be trained on small per-camera RAW sets.
Experimental setup
Datasets: MIT-Adobe FiveK NOD PASCALRAW BDD100K
Benchmarks: FiveK-Nikon FiveK-Canon NOD-Nikon NOD-Sony PASCALRAW
Metrics: PSNR SSIM mAP mAP50 mAP75
Baselines: CycleR2R UNet SRISP InvISP ISPLess RAW-Diffusion RTMDet-s YOLOX-s SSD
Builds on: InvISP ISPLess CycleGAN RAW-Diffusion
Closest papers by meaning
- RAWild: Toward Sensor-Agnostic RAW Object Detection via Physics-Guided Curve and Grid Modeling (NeurIPS 2026 · similarity 0.86)
- Guiding a Harsh-Environments Robust Detector via RAW Data Characteristic Mining (AAAI 2024 · similarity 0.86)
- Towards RAW Object Detection in Diverse Conditions (CVPR 2025 · similarity 0.85)
- RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025 · similarity 0.85)
- RGB-Only Supervised Camera Parameter Optimization in Dynamic Scenes (NeurIPS 2025 · similarity 0.84)
- Logarithmic Lenses: Exploring Log RGB Data for Image Classification (CVPR 2024 · similarity 0.84)
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.