LIFT: Learned Invariant Feature Transform


We introduce a novel Deep Network architecture that implements the full feature point handling pipeline, that is, detection, orientation estimation, and feature description. While previous works have successfully tackled each one of these problems individually, we show how to learn to do all three in a unified manner while preserving end-to-end differentiability. We then demonstrate that our Deep pipeline outperforms state-of-the-art methods on a number of benchmark datasets, without the need of retraining.


K. M. Yi; E. Trulls Fortuny; V. Lepetit; P. Fua : LIFT: Learned Invariant Feature Transform. 2016. European Conference on Computer Vision (ECCV), Amsterdam, The Netherlands, October 8-16, 2016. p. 467-483. DOI : 10.1007/978-3-319-46466-4_28.
J. Auerbach; J. C. Bongard : How robot morphology and training order affect the learning of multiple behaviors. 2009. 2009 IEEE Congress on Evolutionary Computation (CEC), Trondheim, Norway, 18-21 05 2009. p. 39-46. DOI : 10.1109/CEC.2009.4982928.

Teaser Video

This teaser video shows feature matching results with our integrated LIFT pipeline and SIFT, for selected sequences of all three datasets, Strecha, DTU, and Webcam. Our results are significantly better overall compared to SIFT. Note that, in our experiments, SIFT still gives results that are on par with the state-of-the-art when evaluated as a whole pipeline. Please see the paper for details.

Supplementary material

Click the following link for the supplementary appendix for implementation details.