Learning to See in the Dark and Beyond

This thesis presents the design and deployment of a first-of-a-kind perception system capable of functioning off-road 24/7, utilizing multi-spectral inputs following CI/CD approaches. We lead the effort in extending object detection and semantic segmentation to function beyond daytime conditions.

We introduce a new framework for Domain Adaptation (DA), showcased on semantic segmentation, which outperforms existing methods by +40% mIoU in unsupervised scenarios and +25-35% mIoU in semi-supervised scenarios.