Place: Large Lecture Room – CVC
Affiliation: Pattern Recognition Lab Computer Science Department, University of Erlangen-Nuremberg. Germany
During the last 15 years there has been a large body of work on developing robust methods for color constancy. Many of the proposed techniques have focused on relaxing constraining assumptions. The benchmark databases have become increasingly more challenging. Yet, despite all this progress, color constancy techniques have not become widely applicable, despite their potential in improving the accuracy of vision applications such as color segmentation, tracking etc.
Our evaluation of existing methods shows that most color constancy algorithms do not explicitly address key aspects (like camera non-linearities) in the image formation process. Overlooking such issues can have a significant impact on the consistent performance (and thus the reliability) of illuminant–color-estimation algorithms.
The first part of this talk will focus on how an established method can be expanded to become more robust to arbitrary scene analysis. The second part will demonstrate the impact of various factors of the image formation process on the stability of illuminant-estimation methods.