SVDLUT

https://github.com/wontaeaekim/svdlut
Admin
SVDLUT is a lightweight and fast framework for real-time image enhancement that decomposes...
Tokens:1,448
Snippets:19
Trust Score:4.7
License:Apache-2.0
Update:2 months ago
Show doc for...
Tokens:
Raw
### Install Dependencies

Source: https://github.com/wontaeaekim/svdlut/blob/main/README.md

Command to install required Python packages for the model.

```shell
pip install scipy kornia tqdm opencv-python pillow
```

--------------------------------

### Compile Extension

Source: https://github.com/wontaeaekim/svdlut/blob/main/README.md

Command to compile and install the C++ extension for LUT transform and bilateral grid slicing.

```shell
python ./kernel_code/bilateral_slicing_LUTTransform/setup.py install
```

--------------------------------

### Run Demonstration

Source: https://github.com/wontaeaekim/svdlut/blob/main/README.md

Execute the demo script with default or custom paths for pretrained models and images.

```shell
python demo.py
```

```shell
python demo.py --pretrained_path=[pretrained model path] --input_path=[input image path] --output_path=[output image path]
```

--------------------------------

### Evaluate Performance

Source: https://github.com/wontaeaekim/svdlut/blob/main/README.md

Generate test images and calculate metrics using provided Matlab scripts.

```shell
python evaluate.py --pretrained_path=[path of pretrained model] --dataset_name=[fivek or ppr10k] --input_color_space=[sRGB or XYZ]
```

```matlab
(matlab) > ./utils/fivek_calculate_metrics.m [evaluate image dir] [GT dir]
```

```matlab
(matlab) > ./utils/ppr10k_calculate_metrics.m [evaluate image dir] [GT dir] [mask dir]
```

--------------------------------

### Project Directory Structure

Source: https://github.com/wontaeaekim/svdlut/blob/main/README.md

Overview of the project file organization, including source code, datasets, and extension interfaces.

```text
dataset/fiveK
       /ppr10k
project/LUTwithBGrid
       demo_img/                      # image for demo
       kerenel_code/                  # extension codes for lut transform and slicing
       lpips/
       pretrained/                    # pretrainded models
       utils/                         # matlab files for evaluation etc.
       cpp_ext_interface.py           # python interfaces for extension code
       datasets.py                    # codes for dataset classes
       demo.py                        # simple demostration
       evaluate.py                    # evaluation code
       model_losses.py                # losses for our model
       models.py                      # implementaion code for our model
       torchvision_x_functional.py
       train.py                       # train code
```

--------------------------------

### Train Models

Source: https://github.com/wontaeaekim/svdlut/blob/main/README.md

Train the SVDLUT model using various dataset configurations and color spaces.

```shell
python train.py --dataset_name=[fiveK of ppr10k] --input_color_space=[sRGB or XYZ] --batch_size=[batch size]
```

```shell
python train.py --dataset_name=fiveK --input_color_space=sRGB
```

```shell
python train.py --dataset_name=fiveK --input_color_space=XYZ
```

```shell
python train.py --dataset_name=ppr10k --input_color_space=sRGB --batch_size=16 --lut_n_ranks=10 --lut_weight_ranks=10
```

--------------------------------

### Dataset Directory Structure

Source: https://github.com/wontaeaekim/svdlut/blob/main/README.md

Organization of the FiveK and PPR10K datasets used for training and evaluation.

```text
dataset/fiveK
        input/
             JPG/480p/                 # sRGB inputs
             PNG/480p_16bits_XYZ_WB/   # XYZ inputs
        expertC/JPG/480p/              # GT
        train.txt                      
        test.txt

       /ppr10k
        train/
             masks/                    # masks for train
             source/                   # sRGB inputs for train
             source_aug/               # sRGB inputs with 5 versions of augmented images for train 
             target_a/                 # GT of expert a for train
             target_b/                 # GT of expert b for train
             target_c/                 # GT of expert c for train
        val/
             masks/                    # masks for test
             source/                   # sRGB inputs for test
             target_a/                 # GT of expert a for test
             target_b/                 # GT of expert b for test
             target_c/                 # GT of expert c for test
```

--------------------------------

### Cite SVDLUT in BibTeX

Source: https://github.com/wontaeaekim/svdlut/blob/main/README.md

Use this BibTeX entry to cite the SVDLUT research paper in academic publications.

```bibtex
@inproceedings{kim2025SVDLUT,
  title={Lightweight and Fast Real-time Image Enhancement via Decomposition of the Spatial-aware Lookup Tables},
  author={Kim, Wontae, Lee, Keuntek, and Cho, Nam Ik},
  booktitle={International Conference on Computer Vision},
  year={2025}
}
```
Tokens:
Raw