Skip to content

Latest commit

 

History

History
71 lines (43 loc) · 3.5 KB

UltraLED.md

File metadata and controls

71 lines (43 loc) · 3.5 KB

Environment

We employed the same environment as LED, except for the rawpy library. For metric testing in our paper, we used the official rawpy library version 0.16.0 to ensure a fairer comparison against RGB-based methods.

Using the rawpy library improved by the ELD authors—which matches LED—yields better alignment and visualization results, consequently achieving different PSNR/SSIM/LPIPS metrics:

Ratio ✖️50 ✖️100 ✖️200
LED 29.629/0.926/0.073 29.611/0.917/0.101 29.103/0.891/0.152
ELD 29.711/0.925/0.072 29.703/0.915/0.099 29.011/0.887/0.159
PG 28.742/0.921/0.091 28.322/0.903/0.132 27.254/0.855/0.208
UltraLED 31.271/0.939/0.066 31.126/0.929/0.094 30.351/0.898/0.139

Inference

  1. Download the test data and ground truth.
  2. Download the pre-trained ratio map estimator model and the pre-trained raw denoiser model.
python scripts/image_process_ultraled.py 

optional arguments:
  -p the pretrained ratio map estimator network path.
  -pd the pretrained network path for denoising.
  --data_path the folder containing raw images to be processed.
  --save_path output folder for processed images.
  -opt ratio map estimator network architecture options.
  -optd denoising network architecture options.
  --ratio maximum exposure gain ratio.
  [--target_exposure target exposure (overrides ratio).]
  [--bps output bit depth.]

Evaluation

Replace folder_A and folder_B in scripts/test_metrics_ultraled.py with the paths to the test data and ground truth, respectively.

python scripts/test_metrics_ultraled.py

Train

Preparing Training Data

We used the HDR subset from the RAWHDR dataset as our training data. We have converted this data into npz files following the method described in LED. You can directly download the processed npz training data using Baidu Netdisk or Google Drive (note: we only used the HDR data, but we converted all data into npz format and provide their corresponding paired txt files).

To begin training, replace the datasets/train/dataroot parameter in options/base/dataset/pretrain/UltraLED_trainS1.yaml and options/base/dataset/pretrain/UltraLED_trainS2.yaml with the path to the downloaded npz data.

Step1

python led/train.py -opt options/UltraLED/train_step1.yaml

Step2

Replace the network_d_path parameter in options/UltraLED/train_step2.yaml with the path to the model trained in Step 1.

python led/train.py -opt options/UltraLED/train_step2.yaml