Improving interest point object recognition

Vision is a fundamental ability for humans. It is essential to a wide range of activities. The ability to see underpins almost all tasks of our day to day life. It is also an ability exercised by people almost effortlessly. Yet, in spite of this it is an ability that is still poorly understood, and has been possible to reproduce in machines only to a very limited degree. This work grows out of a belief that substantial progress is currently being made in understanding visual recognition processes. Advances in algorithms and computer power have recently resulted in clear and measurable progress in recognition performance. Many of the key advances in recognizing objects have related to recognition of key points or interest points. Such image primitives now underpin a wide array of tasks in computer vision such as object recognition, structure from motion, navigation. The object of this thesis is to find ways to improve the performance of such interest point methods. The most popular interest point methods such as SIFT (Scale Invariant Feature Transform) consist of a descriptor, a feature detector and a standard distance metric. This thesis outlines methods whereby all of these elements can be varied to deliver higher performance in some situations. SIFT is a performance standard to which we often refer herein. Typically, the standard Euclidean distance metric is used as a distance measure with interest points. This metric fails to take account of the specific geometric nature of the information in the descriptor vector. By varying this distance measure in a way that accounts for its geometry we show that performance improvements can be obtained. We investigate whether this can be done in an effective and computationally efficient way. Use of sparse detectors or feature points is a mainstay of current interest point methods. Yet such an approach is questionable for class recognition since the most discriminative points may not be selected by the detector. We therefore develop a dense interest point method, whereby interest points are calculated at every point. This requires a low dimensional descriptor to be computationally feasible. Also, we use aggressive approximate nearest neighbour methods. These dense features can be used for both point matching and class recognition, and we provide experimental results for each. These results show that it is competitive with, and in some cases superior to, traditional interest point methods. Having formed dense descriptors, we then have a multi-dimensional quantity at every point. Each of these can be regarded as a new image and descriptors can be applied to them again. Thus we have higher level descriptors – ‘descriptors upon descriptors’. Experimental results are obtained demonstrating that this provides an improvement to matching performance. Standard image databases are used for experiments. The application of these methods to several tasks, such as navigation (or structure from motion) and object class recognition is discussed.