Stochastic Systems Group
Home Research Group Members Programs  
Demos Calendar Publications Mission Statement Alumni

SSG Seminar Abstract


Capturing the Shape of Objects and Their Variations via the Logarithm of Odds

Kilian Pohl
CSAIL, MIT



The Logarithm of the Odds (LogOdds) is frequently used in areas such as artificial neural networks, economics, and biology. In this talk, we discuss LogOdds as a shape representation that addresses certain computer vision problems. For example, LogOdds embed the non-linear space of signed distance maps within a vector space.

LogOdds is a representation describing the intrinsic knowledge of an object boundary. It therefore encodes the shape of a single object as well as similarities across a group of objects. Furthermore, the representation provides a probabilistic interpretation of shapes within a vector space. This property is useful for non-convex interpolations between space- conditioned-probabilities capturing time-related alterations of objects. We test our representation by incorporating it into an algorithm targeted towards the automatic segmentation of medical images. The Bayesian classification model with our new representation achieves a higher average score than alternative shape models on the test data set.



Problems with this site should be emailed to jonesb@mit.edu