Biologically Motivated Models for Low- and Mid-Level Vision
During my stay at the Odyssee team of INRIA I studied biological models of vision, and tried to see how they relate to the variational approach to vision. Specifically, I studied the FACADE model proposed by S.Grossberg, and explored its relation to more commonly used biological and computer vision models.
Simplifying certain of its stages, and introducing recurrent dynamics allowed the construction of a Lyapunov energy that interprets the models function as minimizing a functional containing both reactive and diffusive terms.
Extending this work, learning such a model is pursued using ground truth edge detection results. For this we build on the variational approach to statistical inference and interpret the network state evolution as estimating the mode of distribution. The learned model's performance is superior to that of standard edge detection algorithms, while the same approach could be extended to a broader range of low- and mid- level tasks.
Results on the Berkeley Benchmark
A brief presentation of this work can be found here
Related Publications
I. Kokkinos, R. Deriche, O. Faugeras and P. Maragos,
Computational Analysis and Learning for a Biologically Motivated Model of Boundary Detection.Accepted in Neurocomputing.
I. Kokkinos, R. Deriche, P. Maragos and O. Faugeras,
A Biologically Motivated and Computationally Tractable Model of Low- and Mid- Level Vision Tasks.,
Proc. European Conference on Computer Vision (ECCV), 2004.
I.Kokkinos, R.Deriche, Olivier Faugeras and P.Maragos,
Towards Bridging the Gap Between Biological and Computational Segmentation.
INRIA Research Report RR-6317