Object category recognition with large training sets

点击翻译按钮获取中文摘要

An important long-term goal for the field of computer vision is to enable a computer to recognize broad categories of objects (for example bicycles, cows, or people) whenever they appear in images. While this seems effortless for a person, the current performance of computer vision systems remains far below that of human vision. The solution to this problem is important for numerous applications, such as automobile driver assistance, image search on the web, or aids to the visually impaired. We believe that a major reason for the limited performance of existing systems is that they use inadequate training data. This research will develop approaches for working with far larger training sets of images, such as tens of thousands of images per image category rather than the few hundred currently used. This will require new learning methods and fast algorithms for scaling up to large amounts of data. We will achieve other improvements in scalability by developing new approaches to feature selection, in which only a small subset of the best potential visual features are selected for each particular discrimination task. Our research group has already developed some widely used software for speeding up similar problems, such as the Fast Library for Approximate Nearest Neighbours (FLANN). We intend to build on our expertise in these areas to exploit large sets of diverse local features and scale training set sizes by at least two orders of magnitude over current capabilities. Preliminary experiments indicate that this will give a large improvement in the accuracy of recognition and enable many new applications of importance to society.