Organizing Flickr: Object Mining using Particular Object Retrieval

Опубликовано 6 сентября 2016, 18:49

The first half of this talk will discuss scalable methods for performing particular object search via visual queries in very large datasets of unordered images. We show how the combination of large vocabularies, fast spatial verification, query expansion and soft assignment can dramatically boost the accuracy of object retrieval to give a fast and reliable system. The second half will explore object mining in image datasets, where the aim is to automatically discover and group images containing the same object. Two methods are examined for performing this mining. The first uses object search to construct a matching graph, where the strength of the edges depends on the spatial consistency between image pairs. Standard clustering techniques are then used to mine for the objects. The second overloads the LDA topic model to represent each image as a mixture of particular objects, with the topic representing both the identity and the spatial layout of visual words. These methods will be demonstrated on several large datasets of Flickr images.