Shu Kong, Charless Fowlkes
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
Consider the image patches at right. Though they lack global context, it is often clear which contain object boundaries and which do not. We are developing algorithms that integrate local image measurements in order to provide global explanations of scenes in terms of objects, surfaces and their spatial relations. This includes algorithms for bottom-up boundary detection and segmentation from local features as well integrating top-down knowledge such as the detection of familiar objects.