Category-sensitive hashing and bloom filter based descriptors for online keypoint recognition
2015 (English)In: 19th Scandinavian Conference on Image Analysis, SCIA 2015, Springer, 2015, 329-340 p.Conference paper (Refereed)Text
In this paper we propose a method for learning a categorysensitive hash function (i.e. a hash function that tends to map inputs from the same category to the same hash bucket) and a feature descriptor based on the Bloom filter. Category-sensitive hash functions are robust to intra-category variation. In this paper we use them to produce descriptors that are invariant to transformations caused by for example viewpoint changes, lighting variation and deformation. Since the descriptors are based on Bloom filters, they support a ”union” operation. So descriptors of matched features can be aggregated by taking their union.We thus end up with one descriptor per keypoint instead of one descriptor per feature (By keypoint we refer to a world-space reference point and by feature we refer to an image-space interest point. Features are typically observations of keypoints and matched features are observations of the same keypoint). In short, the proposed descriptor has data-defined invariance properties due to the category-sensitive hashing and is aggregatable due to its Bloom filter inheritance. This is useful whenever we require custom invariance properties (e.g. tracking of deformable objects) and/or when we make multiple observations of each keypoint (e.g. tracking, multi-view stereo or visual SLAM).
Place, publisher, year, edition, pages
Springer, 2015. 329-340 p.
Lecture Notes in Computer Science, ISSN 0302-9743 ; 9127
Bloom filter, Feature matching, Feature tracking, Hashing, Keypoint recognition, Data structures, Deformation, Hash functions, Stereo image processing, Bloom filters, Feature-tracking, Keypoint, Image analysis
Computer Vision and Robotics (Autonomous Systems)
IdentifiersURN: urn:nbn:se:kth:diva-181549DOI: 10.1007/978-3-319-19665-7_27ScopusID: 2-s2.0-84947935965ISBN: 9783319196640OAI: oai:DiVA.org:kth-181549DiVA: diva2:902628
15 June 2015 through 17 June 2015
QC 201602112016-02-112016-02-022016-02-11Bibliographically approved