Stage de Mastere: Model-based clustering applied to motion segmentation and multi-body factorization

Master thesis

One challenging problem in computer vision is to segment a scene into several moving rigid objects, using as few prior information as possible. Among others, factorization methods have proved to be the method of choice for performing this task. The factorization itself is handled via various versions of robust principal component analysis, such as the GPCA method proposed in Vidal et al., Generalized Principal Component Analysis, IEEE Proceedings of CVPR, 2003.

In this project we propose to take a different approach based on the combination of multi-body factorization methods with unsupervised clustering methods. Within such a framework, each moving object may well be viewed as a cluster. It is suggested to use the machinery of maximum likelihood for mixture of normal distributions. Within such a framework, multi-body factorization can be implemented via the EM algorithm. Each density in the mixture is parameterized by the rigid motion parameters, and hence EM can be briefly paraphrased as follows: assign image features to rigid objects (E) and minimize over the shape and motion parameters of each object (M). This method has successfully been applied in conjunction with 3-D reconstruction using factorization and outlier rejection.

Eligibility: Master student with competences in computer vision, matrix algebra, and probability theory.

Start date: 1 March 2008

Contact person: Radu Patrice HORAUD

Deadline: 1 December 2007