Robust Feature Bundling

Abstract

Abstract. In this work we present a feature bundling technique that aggregates individual local features with features from their spatial neighborhood into bundles. The resulting bundles carry more information of the underlying image content than single visual words. As in practice an exact search for such bundles is infeasible, we employ a robust approximate similarity search with min-hashing in order to retrieve images containing similar bundles. We demonstrate the benefits of these bundles for small object retrieval, i.e. logo recognition, and generic image retrieval. Multiple bundling strategies are explored and thoroughly evaluated on three different datasets.

Bundle min-hashing for logo recognition

“…The authors show that PmH is significantly faster than mH and has identical collision probabilities for sketches as mH in the worst case, but theoretically better recall and precision if the duplicate image region only covers a small area. However, in [15] we observed that PmH performs worse than mH on the logo dataset.…”

“…In [15] we have already shown that IDF-weighting is always beneficial in the bag-of-words framework, even for large vocabularies greater than 1 million words. Thus tf-idf weighting was used in all cases.…”

“…• We adopt the retrieval technique of [15] based on feature bundles to the problem of logo recognition and show that such a system significantly outperforms the current state-of-the-art.…”

Bundle min-Hashing

“…3. In [22] we have already shown that IDF-weighting is always beneficial in the bagof-words framework, even for large vocabularies greater than 1 million words. Thus, tf-idf weighting was used in all cases.…”

“…Inspired by this observation we present a feature bundling technique [22] that builds on visual words. It does not describe each visual word individually but rather aggregates spatial neighboring visual words into feature bundles.…”