Place recognition is an essential component of any Simultaneous Localizationand Mapping (SLAM) system. Correct place recognition is a difficult perceptiontask in cases where there is significant appearance change as the same placemight look very different in the morning and at night or over differentseasons. This work addresses place recognition using a two-step (generative anddiscriminative) approach. Using a pair of coupled Generative AdversarialNetworks (GANs), we show that it is possible to generate the appearance of onedomain (such as summer) from another (such as winter) without needing image toimage correspondences. We identify these relationships considering sets ofimages in the two domains without knowing the instance-to-instancecorrespondence. In the process, we learn meaningful feature spaces, thedistances in which can be used for the task of place recognition. Experimentsshow that learned feature correspond to visual space and can be effectivelyused for place recognition across seasons.
展开▼
