Search

Segmentation is a fundamental and core problem in computer vision research which has applications in many tasks, such as object recognition, content-based image retrieval, and semantic labelling. To partition the data into groups coherent in one or more characteristics such as semantic classes, is often a first step towards understanding the content of data. As information in the real world is generally perceived in multiple modalities, segmentation performed on multi-modal data for extracting the latent structure usually encounters a challenge: how to combine features from multiple modalities and resolve accidental ambiguities. This thesis tackles three main axes of multi-modal segmentation problems: video segmentation and object discovery, activity segmentation and discovery, and segmentation in 3D data. For the first two axes, we introduce non-parametric Bayesian approaches for segmenting multi-modal data collections, including groups of videos and context sensor streams. The proposed method shows benefits on: integrating multiple features and data dependencies in a probabilistic formulation, inferring the number of clusters from data and hierarchical semantic partitions, as well as resolving ambiguities by joint segmentation across videos or streams. The third axis focuses on the robust use of 3D information for various applications, as 3D perception provides richer geometric structure and holistic observation of the visual scene. The studies covered in this thesis for utilizing various types of 3D data include: 3D object segmentation based on Kinect depth sensing improved by cross-modal stereo, matching 3D CAD models to objects on 2D image plane by exploiting the differentiability of the HOG descriptor, segmenting stereo videos based on adaptive ensemble models, and fusing 2D object detectors with 3D context information for an augmented reality application scenario.