作者: Yong Jae Lee , Kristen Grauman
DOI: 10.1007/S11263-009-0252-Y
关键词:
摘要: We present a method to automatically discover meaningful features in unlabeled image collections. Each is decomposed into semi-local that describe neighborhood appearance and geometry. The goal determine for each which of these parts are most relevant, given the content remainder collection. Our first computes an initial image-level grouping based on feature correspondences, then iteratively refines cluster assignments evolving intra-cluster pattern local matches. As result, significance attributed influences image's membership, while related images affect estimated their features. show this mutual reinforcement object-level feature-level similarity improves unsupervised clustering, apply technique categories foreground regions from benchmark datasets.