Object-features for localization under extreme scale changes
Holliday, Andrew · container not stated · 2018
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
· OpenAlex · dissertation · 2018
Abstract
This thesis proposes a technique for the identification and matching of objects between images under significant changes in scale, and the use of these objects as landmarks for scale-robust visual localization. Visual localization under large changes in scale is an important capability in some robotic mapping applications, such as the mapping of coral reefs with a autonomous surface vehicle so that an underwater vehicle can later use the map to localize during missions. Existing approaches are robust only up to a 3x difference in scale between map and query images.
In this thesis, objects are detected in a general, non-semantic fashion, and matchable descriptors are computed from the intermediate activations of a variety of convolutional neural networks. Point-feature matching is then performed only between the image regions corresponding to matched objects. Experiments are conducted in simulation and in real-world outdoor scenes, exhibiting up to a 7x change in scale.
In these experiments, three major variations of the proposed localization scheme are considered: learned object proposals with descriptors sampled from sub-regions of a neural net layer, Selective Search object proposals with the same sub-region descriptors, and Selective Search object proposals with the full network layer activations used as descriptors. All are compared against state-of-the-art SIFT features. The best-performing variation is shown to provide superior localization performance to SIFT under changes in scale of 3x and greater.
Open the resolved locator · 10.82308/27872
Metadata licence: OpenAlex metadata (CC0)