Student research opportunities
Multiscale Conditional Random Fields for Scene Recognition
Project Code: CECS_874
This project is available at the following levels:
CS single semester, Honours, Summer Scholar, Masters
Keywords:
Machine Learning, Computer Vision, Graphical Models
Supervisor:
Dr Justin DomkeOutline:
There is great interest recently in the problem of scene segmentation: given an input image, predict the class of each pixel, for example, road, sky, or grass. A very successful approach for this is a conditional random field (CRF). A CRF defines a probability distribution over segmentations, ideally giving higher probability to true segmentations. Sophisticated algorithm have been developed to perform inference and learning of these models.
At the same time, a decades-old principle in computer vision is scale-invariance. Roughly speaking, if we take a picture of a scene with two different resolutions, our algorithm should give approximately the same answer.
CRFs, as they are normally used, however, are quite dependent on the particular scale at which they are run.
Goals of this project
In this project, you will design and implement a CRF model that explicitly enforces scale-invariance. Off-the-shelf software tools can mostly be used for this, though you will need to get your hands dirty deriving image features, and processing results. The hope is that, by enforcing scale-invariance, a CRF model will show an improved ability to generalize to new data. More senior students should also develop the theory for their model-- such as giving bounds on how close it is to scale invariance, or generalization guarantees.
Requirements/Prerequisites
Familiarity with linear algebra, probability, and image processing. Some experience with numerical computing in a language such as Matlab, Octave, Mathematica, R, Python/NumPy, IDL, Fortran, or C++/Eigen. You will need to learn about probabilistic graphical models.
Background Literature
Conditional Random Fields in General:
http://homepages.inf.ed.ac.uk/csutton/publications/crftut-fnt.pdf
Scene segmentation using CRFs:
eprints.pascal-network.org/archive/00003655/01/VT07a.pdf
http://phd.gccis.rit.edu/justindomke/papers/2013pami.pdf
ai.stanford.edu/~koller/Papers/Gould+al:ICCV09.pdf
Scale Space and Scale Invariance:
www.stat.cmu.edu/~annlee/IJCV01_occlusions.pdf
http://en.wikipedia.org/wiki/Scale_space
