Generating svBRDF textures from photos taken with mobile phone. [arxiv]
The pipeline is implemented in Python 3 with the following packages as additional dependencies:
- Numpy
- Scipy
- matplotlib
- sklearn
- cv2
- exifread
Anaconda distribution is recommended for simplifying the setup of the environment.
The pipeline requires a minimum of two input photos: one under the ambient illumination and one under the flash light. Then place them under the /photos and create a JSON file describing the input images. There is one example photos/green_towel_low, which you can refer to. You should preserve the EXIF of the images, since information like exposure time and focal length is required.
For each camera, you also need to shoot a photo of a 18% gray card to decide the intensity of the flash light, and a series of photos of a generated color palette (./photos/colorcard.jpg) to recover the camera's response curve. These are described as calibration images. Include them in the [list_json_file] (like ./photos/list_low.json) as well as all the folders of the materials you would like to process.
Prepare all the required input photos as above, and run
python ./scripts/fit_all.py ./photos/[list_json_file]For the example contained in the repository
python ./scripts/fit_all.py ./photos/list_low.json
And svBRDF textures will be generated in ./photos/[material_name]/out/
A Python script ./blender/SetupScene.py can be used to quickly import all the textures into a 3D scene ./blender/material.blend. Tested on Blender 2.80.