MatchMiner: Efficient Spanning Structure Mining in Large Image Collections

摘要: Many new computer vision applications are utilizing large-scale data- sets of places derived from the many billions photos on Web. Such often require knowledge visual connectivity structure these image collections--describing which images overlap or otherwise related--and an important step in understanding this is to identify connected components underlying graph. As graph initially unknown, problem can be posed as one exploring between quickly possible, by intelligently selecting a subset pairs for feature matching and geometric verification, without having test all O(n2) possible pairs. We propose novel, scalable algorithm called MatchMiner that efficiently explores relations images, incorporating ideas relevance feedback improve decision making over time, well simple yet effective rank distance measure detecting outlier images. Using ideas, our automatically prioritizes potentially connect contribute large components, using information-theoretic decide next. Our experimental results show find collections, significantly outperforming state-of-the-art methods.