Object Recognition via Local Patch Labelling

Abstract

In recent years the problem of object recognition has received considerable attention from both the machine learning and computer vision communities. The key challenge of this problem is to be able to recognize any member of a category of objects in spite of wide variations in visual appearance due to variations in the form and colour of the object, occlusions, geometrical transformations (such as scaling and rotation), changes in illumination, and potentially non-rigid deformations of the object itself. In this paper we focus on the detection of objects within images by combining information from a large number of small regions, or ‘patches’, of the image. Since detailed hand-segmentation and labelling of images is very labour intensive, we make use of ‘weakly labelled’ data in which the training images are labelled only according to the presence or absence of each category of object. A major challenge presented by this problem is that the foreground object is accompanied by widely varying background clutter, and the system must learn to distinguish the foreground from the background without the aid of labelled data. In this paper we first show that patches which are highly relevant for the object discrimination problem can be selected automatically from a large dictionary of candidate patches during learning, and that this leads to improved classification compared to direct use of the full dictionary. We then explore alternative techniques which are able to provide labels for the individual patches, as well as for the image as a whole, so that each patch is identified as belonging to one of the object categories or to the background class. This provides a rough indication of the location of the object or objects within the image. Again these individual patch labels must be learned on the basis only of overall image class labels. We develop two such approaches, one discriminative and one generative, and compare their performance both in terms of patch labelling and image labelling. Our results show that good classification performance can be obtained on challenging data sets using only weak training labels, and they also highlight some of the relative merits of discriminative and generative approaches.