Object Discovery and Cosegmentation Based on Dense Correspondences.

We propose to do object discovery and cosegmentation in noisy datasets with utilization of CNN features. We use an object discovery framework which supposes that common object patterns are sparse concerning transformations across images. The key issue is then how to take advantage of the interrelations among images. Since an image normally matches better with similar images containing the same object than noise images, we exploit the image matching situations of a dataset to capture the interrelations information in it. Comparing with local feature matching, we aim to estimate the dense correspondences between regions with common semantics using mid-level visual information, which captures the visual variability within the whole dataset. Besides, due to the powerful feature learning ability of deep models, we adopt VGG features to do unsupervised clustering and find representative candidates as a prior knowledge. Experiments on noisy datasets show the effectiveness of our method.