From local features to local regions

• Authors: Stefan Romberg
• Venue: MM '11 Proceedings of the 19th ACM international conference on Multimedia
• Year: 2011

Local features are omnipresent in computer vision applications and an important building block for applications like object recognition and image retrieval. Such applications often involve feature matching or use an inverted index to efficiently retrieve similar local features for a given query. In such cases it is commonly known that local features are often not informative yielding mismatches and false positives when used for feature matching (see Figure 1)or retrieval. As a result retrieval systems employ expensive post-retrieval verification steps. These observations underline the importance of embedding spatial information of a local image region directly within the index to decrease the number of false positives upon retrieval. We present our latest method for embedding spatial information into an index by bundling local feature triples for logo recognition. We give an overview of our feature representation, describe currently ongoing research and pose open questions for discussion.