Detecting partially occluded objects in images

Object detection is one of the oldest problems in computer vision which have eluded a ’grand unified theory’ till date. Occlusion of objects is a major concern for many object detection algorithms, and indeed all algorithms that output a simple bounding box. The first part of this thesis explores existing research in segmentation-aware object detection, which depends on pixel level binary object/notobject segmentation of the contents of the bounding box for detection. This incorporation of pixellevel labelling provides an intuitive method of dealing with occlusion. The labelling is achieved by performing maximum a-posteriori inference on a grid binary Markov Random Field (MRF) defined over the bounding box. The second part of this thesis describes an effective approach for defining pairwise terms in the MRF that encourages visual consistency and a strategy using this framework for detecting pairs of objects localizing them and labelling the contents of their bounding boxes such that the ‘agreement’ between the two objects is maximized.