Identifying humpback whale flukes by sequence matching of trailing edge curvature
Loading...
Authors
Jablons, Zachary
Issue Date
2016-05
Type
Electronic thesis
Thesis
Thesis
Language
ENG
Keywords
Computer science
Alternative Title
Abstract
To our knowledge, this is the first method that can achieve this level of accuracy on humpback fluke identification without extensive manual effort at inference time.
Photographic identification of humpback whale (Megaptera novaeangliae) flukes (i.e. their tail) is an important task in marine ecology, and is used in tracking migration patterns and estimating populations. In this thesis, we lay out a method that automates the photo-identification of humpback flukes using the "trailing edge" of the fluke. The method uses convolutional networks to identify keypoints on the fluke and possible trailing edge locations. It then uses this information to extract a detailed trailing edge and its curvature, which is then matched to other trailing edges in the database via dynamic time warping. Using this method, we achieve nearly 80% top-1 ranking accuracy on a large subset of the SPLASH dataset consisting of about 400 identified individuals. We also show that in combination with HotSpotter, a general pattern based matching algorithm, we can achieve 93% accuracy.
Photographic identification of humpback whale (Megaptera novaeangliae) flukes (i.e. their tail) is an important task in marine ecology, and is used in tracking migration patterns and estimating populations. In this thesis, we lay out a method that automates the photo-identification of humpback flukes using the "trailing edge" of the fluke. The method uses convolutional networks to identify keypoints on the fluke and possible trailing edge locations. It then uses this information to extract a detailed trailing edge and its curvature, which is then matched to other trailing edges in the database via dynamic time warping. Using this method, we achieve nearly 80% top-1 ranking accuracy on a large subset of the SPLASH dataset consisting of about 400 identified individuals. We also show that in combination with HotSpotter, a general pattern based matching algorithm, we can achieve 93% accuracy.
Description
May 2016
School of Science
School of Science
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY