Holistic shape-based object recognition using bottom-up image structures

Object recognition performance that rivals human ability is one of the primary goals of computer vision research. While recognition may take many forms, key tasks include detection, estimating object pose, and segmenting the object from the background. This thesis explores the use of holistic shape matching for recognition using bottom-up image structures such as image segments and contours for all of these tasks. Holistic shape matching utilizes global information about object shape for matching, rather than local image features which often contain too little information to match reliably to the object model. By examining different tasks related to object recognition, we demonstrate the value of holistic shape matching in a broad range of problems, including perceptual grouping, human pose estimation, and object recognition. First, we introduce a method for perceptual grouping of contours in an image into larger groups that uses holistic shape matching to estimate the motion of image contours to a second, related image and group them according to similarties in motion. Holistic shape matching provides scoring for different motion hypotheses, and the final grouping is achieved using a min-cut graph cut to infer the cluster assignment for each contour. Secondly, we describe a method for human pose estimation using image segments that incrementally merges segments into hypotheses for increasingly larger regions of the human body. These hypotheses are verified by matching against a set of shape exemplars using a shape matching method that is articulation-invariant and incorporates holistic shape information. Lastly, we present a two-step method for automatically learning an object detector for an object category from positive and negative images annotated with bounding boxes. In the first step, the object shape is learned from bottom-up image contours extracted in the positive images by searching for lucky contours that can explain large portions of the shape of positive examples. Given the learned shape, the second step trains a discriminative object detector that matches the shape against contours, emphasizing shape features that provide good detection performance. We compare against baselines and previous work that do not use holistic evaluation of shape features to demonstrate its value