We proposes a novel $\bf{\rho}$-Vision to directly perform high-level semantic understanding and low-level compression using RAW images. The framework is demonstrated to provide better detection accuracy and compression than RGB-domain counterparts and is shown to be able to generalize across different camera sensors and task-specific models. Additionally, it has the potential to reduce ISP computation and processing time.
In this repo, we release our Unpaired CycleR2R code and pretrained models. With Unpaired CycleR2R, you could train your RAW model with diversity and realistic simulated RAW images and then deploy them in the real-world directly.
pip install -r requirements.txt
(required) Download the MulitRAW LUCID subset (passwd: x2un).
(optional) Download the BDD100K.
(optional) Download the Cityscapes.
(optional) Download the Flicker2W.
The datasets folder will be like:
datasets
├─multiRAW
│ ├─iphone_xsmax
│ ├─huawei_p30pro
│ ├─asi_294mcpro
│ └─oneplus_5t
├─(optional) bdd100k
├─(optional) cityscapes
└─(optional) flicker
Source RGB | Target RAW | Model |
---|---|---|
BDD100K | iPhone XSmax | link |
BDD100K | Huawei P30pro | link |
BDD100K | asi 294mcpro | link |
BDD100K | Oneplus 5t | link |
Cityscapes | iPhone XSmax | link |
Flicker2W | iPhone XSmax | link |
Flicker2W | Huawei P30pro | link |
Flicker2W | asi 294mcpro | link |
Flicker2W | Oneplus 5t | link |
python train.py configs/unpaired_cycler2r/unpaired_cycler2r_in_bdd100k_rgb2iphone_raw_20k.py
Please download the pretrained model first.
You coud inference using command,
python inference.py --ckpt bdd100k_rgb_to_iphone_raw.pth --rgb resources/bdd100k.jpg
or in your code
from inference import DemoModel
ckpt_path = 'bdd100k_rgb_to_iphone_raw.pth'
rgb_path = 'resources/bdd100k.jpg'
model = DemoModel(ckpt_path)
rgb = imread(rgb_path).astype(np.float32) / 255
rgb = torch.from_numpy(rgb).permute(2, 0, 1)[None]
model = model.cuda()
rgb = rgb.cuda()
raw = model(rgb, mosaic=False)
If your find our dataset and work are helpful for your research, please cite our paper.
@article{li2022efficient,
title={Efficient Visual Computing with Camera RAW Snapshots},
author={Zhihao Li, Ming Lu, Xu Zhang, Xin Feng, M. Salman Asif, and Zhan Ma},
journal={arxiv},
url={https://arxiv.org/pdf/2212.07778.pdf},
year={2022},
}