Team: play1play

for 'Night Photography Rendering Challenge'

How to run

Install python >= 3.9
Install the required packages: pip install -r requirements.txt
Put the test images in the data folder or specify the input image path with -p option.
Run the script:

python run.py -p <input_image_path> -o <output_image_path>

The output images will be saved in the output folder or specifiy the output path with -o option.

docker build -t play1play .

docker run -it --rm --gpus=all -v $(pwd)/data:/data play1play ./run.sh

Do not forget to to --gpus flag, our model requires GPU to run.

2024.3.21:

Final version v3.0 is released for 3rd validation!

Key features:

Utilize the patch-based and calibration-based white-balance algorithm to improve the image quality.
Modify the resizing strategy to improve the image quality.

2024.3.16:

Release v2.0 for 3rd validation!

Key features:

Algorithm changes:

Add luma shading correction (LSC) module
Add auto-contrast module (dynamic gamma)
LSC, LTM, auto-contrast module can dynamiclly adjust the parameters based on the camera gain from the metadata.
Add another white-balance process at the end of the pipeline to further improve the image quality.

Key algorithm parameter explanation:

k_th: defines the threshold for the noise level, higher means tolerant to noise, lower means more sensitive to noise. For this sensor, default is 2.5e-3. Note that this paramerts are shared by all the modules, and hence defined as the member varibale in RawProcessingPipelineDemo class.
s of local_tone_mapping: defines how to apply the gain map to different image channels. Higher means more saturated, lower means less saturated. Default is 0.7. Currently, s is automatically adjusted based on the camera gain from the metadata.

2024.3.7:

release v1.0 for 3rd validation!