Object classes are central to computer vision and have been the focus of substantial research in the last fifteen years. This thesis addresses the tasks of localizing entire objects in images (object class detection) and localizing their semantic parts (part detection). We present four contributions, two for each task. The first two improve existing object class detection techniques by using context and calibration. The other two contributions explore semantic part detection in weakly-supervised settings. First, the thesis presents a technique for predicting properties of objects in an image based on its global appearance only. We demonstrate the method by predicting three properties: aspect of appearance, location in the image and class membership. Overall, the technique makes multi-component object detectors faster and improves their performance. The second contribution is a method for calibrating the popular Ensemble of Exemplar- SVM object detector. Unlike the standard approach, which calibrates each Exemplar- SVM independently, our technique optimizes their joint performance as an ensemble. We devise an efficient optimization algorithm to find the global optimal solution of the calibration problem. This leads to better object detection performance compared to using independent calibration. The third innovation is a technique to train part-based model of object classes using data sourced from the web. We learn rich models incrementally. Our models encompass the appearance of parts and their spatial arrangement on the object, specific to each viewpoint. Importantly, it does not require any part location annotation, which is one of the main limits to training many part detectors. Finally, the last contribution is a study on whether semantic object parts emerge in Convolutional Neural Networks trained for higher-level tasks, such as image classification. While previous efforts studied this matter by visual inspection only, we perform an extensive quantitative analysis based on ground-truth part location annotations. This provides a more conclusive answer to the question.

Domestic robots developed to support human beings by performing daily tasks such as cleaning should also be able to help in emergencies by finding, analysing, and assisting persons in need of first aid. Here such a robot capable of performing some useful task related to first aid is referred to as a First Aid Mobile Robot (FAMR). One challenge which to the author's knowledge has not been solved is how such a FAMR can find a fallen person's pose within an environment, recognising locations of points of interest for first aid such as the mouth, nose, chin, chest and hands on a map. To overcome the challenge, a new approach is introduced based on leveraging a robot's capabilities (multiple sensors and mobility), called AHBL. AHBL comprises four steps: Anomaly detection, Human detection, Body part recognition, and Localisation on a map. It was broken down into four steps for modularity (e.g., a different way of detecting anomalies can be slipped in without changing the other modules) and because it was not clear which step is hardest to implement. As a result of evaluating AHBL, a FAMR developed for this work was able to find the pose of a fallen person (a mannequin) in a known environment with an average success rate of 83%, and an average localisation discrepancy of 1.47cm between estimated body part locations and ground truth. The presented approach can be adapted for use in other robots and contexts, and can act as a starting point toward designing systems for autonomous robotic first aid.

Computational modeling of part relations of shapes is a challenging problem that has been addressed by many researchers since sixties. The most important source of the difficulty is the continuous nature of shapes, which makes the expression of shape very difficult in terms of discrete parts. When discrete parts are combined, they fuse and yield new parts, i.e. parts emerge. There is a number of methods that support emergent part detection. However all of these methods are based on strong assumptions in terms of what constitute a part. There is a need for a generic solution that treats a shape independently of any restriction resulting from analytical, geometrical, or logical abstractions. To this end, we have developed two novel strategies, which can be used both separately and jointly. Both strategies are relatable to the algebraic formalization of shape grammars (by Stiny). In the course of this thesis work, we have introduced a novel data structure called Over-Complete Graph to address the problem of part embedding in the existence of discrete registration marks / and we have developed a novel and robust method for the automatic selection of registration marks. Both developments are certainly useful for other visual problems. On the application side, we have tested our techniques on puzzling Seljuk patterns (from Kayseri) to demonstrate how the developed techniques give way to computational creativity. Apart from the techniques we have developed, the most important contribution of our work is that shapes are treated as perceived wholes rather than composed, as compellingly demonstrated by Seljuk pattern experiments.

No description available.