Fully-trainable deep matching

James Thewlis, Shuai Zheng, Philip Torr and Andrea Vedaldi

Abstract

Deep Matching (DM) is a popular method for quasi-dense image matching due to the quality of the correspondences that it can establish. However, DM, as originally proposed, is not a deep neural network and cannot be trained end-to-end via backpropagation. In this paper, we remove this limitation by rewriting the complete DM algorithm as a convolutional neural network. This results in a novel deep architecture for image matching that involves a number of new layer types and that, similar to recent networks for image segmentation, has a $U$-topology. We demonstrate the utility of the approach by improving the performance of DM by learning it end-to-end on an image matching task.

Session

Video events, robot vision and deep learning

Files

PDF iconExtended Abstract (PDF, 2M)
PDF iconPaper (PDF, 2M)

DOI

10.5244/C.30.145
https://dx.doi.org/10.5244/C.30.145

Citation

James Thewlis, Shuai Zheng, Philip Torr and Andrea Vedaldi. Fully-trainable deep matching. In Richard C. Wilson, Edwin R. Hancock and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 145.1-145.12. BMVA Press, September 2016.

Bibtex

        @inproceedings{BMVC2016_145,
        	title={Fully-trainable deep matching},
        	author={James Thewlis, Shuai Zheng, Philip Torr and Andrea Vedaldi},
        	year={2016},
        	month={September},
        	pages={145.1-145.12},
        	articleno={145},
        	numpages={12},
        	booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
        	publisher={BMVA Press},
        	editor={Richard C. Wilson, Edwin R. Hancock and William A. P. Smith},
        	doi={10.5244/C.30.145},
        	isbn={1-901725-59-6},
        	url={https://dx.doi.org/10.5244/C.30.145}
        }