Basel Illumination Prior 2017
Illumination is a major challenge in computer vision systems. We propose a robust illumination estimation technique and applied it to the Annotated Facial Landmarks in the Wild (AFLW). This database contains face images taken in various settings and under various in- and outdoor illumination conditions. Using this dataset we estimated the first illumination prior for real world images. The prior is a probability distribution of natural illumination conditions. The prior is built from illumination estimations on 14'348 manually selected face images and depicts an empirical real world illumination distribution. The illumination is modeled using spherical harmonics. We approximate the illumination condition using the first three bands. We estimate a multivariate normal distribution to get a parametrized representation. To calculate the eigenmodes we perform principal component analysis. Our prior closes a gap in generative modeling and can be used in a wide range of applications. It can be integrated in probabilistic image analysis frameworks. Further more, the resulting illumination prior can improve discriminative methods which aim to be robust against illumination. This is especially helpful for data-greedy methods like deep learning. Those methods already include a 3DMM as prior for face shape and texture to augment or synthesize training data and could further profit from using the proposed illumination prior. The proposed illumination prior is an ideal companion of the 3DMM and allows to synthesize more realistic images.
Random illumination conditions sampled from our illumination prior.
We provide the raw data of the estimated illumination conditions, as well as software to generate new illumination conditions from the illumination prior. For further details please consult the below mentioned publication.
For scientific usage please cite the following publication and refer to the Basel Illumination Prior 2017 (BIP2017):
- Occlusion-aware 3D Morphable Models and an Illumination Prior for Face Image Analysis [pdf]]
Bernhard Egger, Sandro Schoenborn, Andreas Schneider, Adam Kortylewski, Andreas Morel-Forster, Clemens Blumer and Thomas Vetter
International Journal of Computer Vision, 2018
For other usage please give appropriate credit to the Graphics and Vision Research Group at the University of Basel.
(unpacking under Windows takes a while)