Online Invariance Selection for Local Feature Descriptors
PubDate: July 2020
Teams: ETH Zurich；Microsoft
Writers: Rémi Pautrat Viktor Larsson Martin R. Oswald Marc Pollefeys
To be invariant, or not to be invariant: that is the question formulated inthis work about local descriptors. A limitation of current feature descriptorsis the trade-off between generalization and discriminative power: moreinvariance means less informative descriptors. We propose to overcome thislimitation with a disentanglement of invariance in local descriptors and withan online selection of the most appropriate invariance given the context. Ourframework consists in a joint learning of multiple local descriptors withdifferent levels of invariance and of meta descriptors encoding the regionalvariations of an image. The similarity of these meta descriptors across imagesis used to select the right invariance when matching the local descriptors. Ourapproach, named Local Invariance Selection at Runtime for Descriptors (LISRD),enables descriptors to adapt to adverse changes in images, while remainingdiscriminative when invariance is not required. We demonstrate that our methodcan boost the performance of current descriptors and outperformsstate-of-the-art descriptors in several matching tasks, when evaluated onchallenging datasets with day-night illumination as well as viewpoint changes.