Thesis Defense: Learning Static Object Segmentation from Motion Segmentation

Michael G. Ross , MIT CSAIL
Date: Wednesday, August 10 2005

Dividing an image into its constituent objects can be a useful first step in many visual processing tasks, such as object classification or determining the arrangement of obstacles in an environment. Motion segmentation is a rich source of training data for learning to segment objects by their static image properties. Background subtraction can distinguish between moving objects and their surroundings, and the techniques of statistical machine learning can capture information about objects' shape, size, color, brightness, and texture properties. Presented with a new, static image, the trained model can infer the proper segmentation of the objects present in a scene. The algorithm presented in this work uses the techniques of Markov random field modeling and belief propagation inference, outperforms a standard segmentation algorithm on an object segmentation task, and outperforms a learned boundary detector at determining object boundaries on the test data.