| Summary: | Recognition of categories of objects is one of the central problems of computer vision. The human visual system has an unmatched ability to recognize objects across multiple modalities and appearances. Recognition of objects via their shapes is one of the primary reasons why humans are able to perform well in vision-related tasks. However, automatic algorithms that try to recognize objects are far from such ability. Therefore, this thesis concentrates on developing better representations of shapes to aid in object retrieval and object detection. We begin by questioning the implicit assumption that all of the shape information lies in the contours, and show that making use of the interior properties of the shapes will produce better results while matching shapes. To this end, we introduce a novel descriptor, namely, the Solid Shape Context. We then hypothesize that not all parts of the shape are equally important while extracting the shape properties, and try to identify the most discriminative parts. We extract the most discriminative parts of a shape by comparing each shape to its closest rivals and propose a means to improve the discriminative capability of standard shape descriptors. We then propose a simple and intuitive way to obtain robust neighborhoods, which help in computing the ``true" distances between shapes. This is achieved by mining additional information that was, until now, unidentified. In addition, we provide soft probabilistic measures for the inclusion or removal of a node from a local neighborhood. The ability to measure confidence of nodes being a part of the neighborhood was not explored till now. This work opens many avenues for future research in the rapidly growing field of retrieval. Finally, as an unifying work, we propose a framework for performing shape-based object detection in real-world images, which also allows for the identification of object parts. This work bridges the gap between two independent, but actively researched, threads in the field of object detection. We propose a structured prediction approach for predicting object part labels, where the label of each part gets influenced by its neighboring parts.
|
|---|