Computing image patch descriptors for correspondence problems relies heavily on hand-crafted feature transformations, e.g. SIFT, SURF. In this paper, we explore a Siamese pairing of fully connected neural networks for the purpose of learning discriminative local feature descriptors. Resulting ANN computes 128-D descriptors, and demonstrates consistent speedup as compared to such state-of-the-art methods as SIFT and FREAK on PCs as well as in embedded systems. We use L₂ distance to reflect descriptor similarity during both training and testing. In this way, feature descriptors we propose can be easily compared to their hand-crafted counterparts. We also created a dataset augmented with synthetic data for learning local features, and it is available online. The augmentations provide training data for our descriptors to generalise well against scaling and rotation, shift, Gaussian noise, and illumination changes.

Date Published: 8 February 2017
PDF: 5 pages
Proc. SPIE 10253, 2016 International Conference on Robotics and Machine Vision, 1025306 (8 February 2017); doi: 10.1117/12.2266351

Published in SPIE Proceedings Vol. 10253:
2016 International Conference on Robotics and Machine Vision
Alexander V. Bernstein; Adrian Olaru; Jianhong Zhou, Editor(s)