Unsupervised Robust Feature-Based Partition Ensembling to Discover Categories

Abstract

The design of novel robust image descriptors is still a formidable problem. Different features, with different capabilities, are introduced every year. However, to explore how to combine them is also a fundamental task. This paper proposes two novel strategies for aggregating different featurebased image partitions to tackle the challenging problem of discovering objects in unlabeled image collections. Inspired by consensus clustering models, we introduce the Aggregated Partition (AP) approach, which, starting from a set of weak input partitions, builds a final partition where the disagreements with the input partitions are optimized. We then generalize the AP formulation and derive the Selective AP, which automatically identifies the subset of features and partitions that further improves the precision of the final partition. Experiments on three challenging datasets show how our methods are able to consistently outperform competing methods, reporting state-of-the-art results.