Addressing Challenging Place Recognition Tasks using Generative Adversarial Networks

Place recognition is an essential component of any Simultaneous Localization and Mapping (SLAM) system. Correct place recognition is a difficult perception task in cases where there is significant appearance change as the same place might look very different in the morning and at night or over different seasons. This work addresses place recognition using a two-step (generative and discriminative) approach. Using a pair of coupled Generative Adversarial Networks (GANs), we show that it is possible to generate the appearance of one domain (such as summer) from another (such as winter) without needing image to image correspondences. We identify these relationships considering sets of images in the two domains without knowing the instance-to-instance correspondence. In the process, we learn meaningful feature spaces, the distances in which can be used for the task of place recognition. Experiments show that learned feature correspond to visual space and can be effectively used for place recognition across seasons.