MINIMAL BASIS REPRESENTATION FOR GENERAL MOTION SEGMENTATION

Abstract

While motion segmentation has been an active research area, the model selection aspect has often been neglected. Due to the difficulty of simultaneously estimating the number of motion and segmenting the trajectories, the number of motion is often assumed known.

In this thesis, we present a model selection mechanism based on finding the minimal basis subspace representation. This model selection mechanism is the enabler for our proposed general motion segmentation work that is capable of strong competitive performance for both rigid and non-rigid motion. The good performance can be attributed to the explicit modeling of overlapping subspaces by identifying the shared bases, which is also key to ensuring the recovery of a global shape in non-rigid structure from motion.

We first apply our general motion segmentation work to rigid motion segmentation by evaluating both the model selection and segmentation performance against the state-of-the-art rigid motion segmentation algorithms, using the standard Hopkins 155 and extended Hopkins 380 dataset. These evaluations show that our work offers the best performance.

Based on this general motion segmentation work, we develop a new subspace segmentation approach to non-rigid structure from motion. This new subspace segmentation approach decomposes a complex non-rigid motion into sub-groups of relatively simpler motion, which can be more easily reconstructed. Even without the benefit of ground truth, our approach compares favorably with the state-of-the-art works.