ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge

(27.02.2025) ReRAW was accepted at CVPR2025! 🚀

This repository contains code to train ReRAW to convert RGB images into sensor specific RAW.

Training ReRAW

Prepare a paired dataset of RAW and RGB files, then modify the config at ./config/cfg_example.py.

Training ReRAW involves first preparing a stratified sampling training patch dataset from the full RAW and RGB images.

1. Stratified Sampling

Make sure the dataset and cfg_sample parameters are set to match your dataset. Modify dataset to add the original folder root, RGB images, and RAW images locations. Update the white level and black level of the sensor. In cfg_sample, only the output_folder_root should contain the root where the resulting stratified sampling dataset will be saved.

Then run:

python3 stratified_sampling.py -c ./config/cfg_example.py -n 2

where -n denotes the number of parallel workers.

2. Training

With the rgb-context, rgb-sample and rggb-target folders are populated, run:

python3 train.py -g 0 -c ./config/cfg_example.py

where -g denotes GPU number.

The training will run and save the checkpoint at a timestamp folder in ./outputs/.

Running ReRAW

To run ReRAW and convert a folder of RGB images into sensor-specific RAW, run:

python3 convert.py -g 0 -f ./outputs/1731913987 -i ./example/rgb -o ./example/converted -n 2 -r 1

where:

Parameter	Description
-g	GPU number.
-f	Model output folder.
-i	RGB input folder.
-o	Output RGGB folder.
-n	No of parallel workers.
-r	Convert RGGB to RGB for visualisation.