The project webpage can be found here: https://team.inria.fr/graphdeco/projects/multi-materials/
The data for the pre-training can be found on the project webpage.
Empowered by deep learning, recent methods for material capture can estimate a spatially-varying reflectance from a single photograph. Such lightweight capture is in stark contrast with the tens or hundreds of pictures required by traditional optimization-based approaches. However, a single image is often simply not enough to observe the rich appearance of real-world materials. We present a deep-learning method capable of estimating material appearance from a variable number of uncalibrated and unordered pictures captured with a handheld camera and flash. Thanks to an order-independent fusing layer, this architecture extracts the most useful information from each picture, while benefiting from strong priors learned from data. The method can handle both view and light direction variation without calibration. We show how our method improves its prediction with the number of input pictures, and reaches high quality reconstructions with as little as 1 to 10 images — a sweet spot between existing single-image and complex multi-image approaches.
This code relies on Tensorflow 1.X but can be adapted to TF 2.X with the following compatibility code:
Replace tensorflow import everywhere by:
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
It is based on python 3.X, numpy, imageio and opencv for python.
This network is trained to use 256x256 gamma corrected pictures (assumes gamma 2.2, this is set in dataReader.py) and output both the linear and gamma corrected ("gammad", gamma only applies to the albedos) parameters. Higher resolution tend to work less well despite the convolutional nature of the network (see https://team.inria.fr/graphdeco/projects/large-scale-materials/ supplemental materials).
The model used is the one described in our single image capture paper: https://github.com/valentin-deschaintre/Single-Image-SVBRDF-Capture-rendering-loss (similar to Adobe Substance), changing the rendering model implementation to render the results will cause strong appearance difference as different implementations use the parameters differently (despite sharing their names, for example diffuse and specular will be controled for light conservation or roughness will be squared)!
To retrain the network, the basic version is to run trainNetwork.sh You can find the training dataset (1.3Go) here: https://repo-sam.inria.fr/fungraph/multi_image_materials/supplemental_multi_images/materialsData_multi_image.zip (A higher definition version of this dataset is available in our more recent project repository: https://github.com/valentin-deschaintre/Guided_fine_tuning_SVBRDF)
First download the trained weights here: https://repo-sam.inria.fr/fungraph/multi_image_materials/supplemental_multi_images/checkpointTrained.zip
You can then run testModel.sh
To achieve the best results, one of the inputs should have an approximately centered flash image.
To use a single image without the ground truth concatenated to it, use the --mode eval rather than --mode test.
If you use our code, please cite our paper:
@Article{DADDB19, author = "Deschaintre, Valentin and Aittala, Miika and Durand, Fr'edo and Drettakis, George and Bousseau, Adrien", title = "Flexible SVBRDF Capture with a Multi-Image Deep Network", journal = "Computer Graphics Forum (Proceedings of the Eurographics Symposium on Rendering)", number = "4", volume = "38", month = "July", year = "2019", keywords = "Reflectance modeling, Image processing, Material capture, Appearance capture, SVBRDF, Deep learning", url = "http://www-sop.inria.fr/reves/Basilic/2019/DADDB19" }