Seeing Beyond Seeing with Enhanced Deep Tracking

Deep Tracking ([1]) is a recently proposed theoretical framework for end-to-end unsupervised visual tracking. Researchers have successfully applied the framework to directly predict full unoccluded space occupancy grid map from raw laser input data. Compared to traditional visual tracking methods on robotics and autonomous driving, which heavily relies on real world knowledge, the learning framework demonstrates very good potential without the need of feature engineering. In this work I further investigate the practical limitation of the current Deep Tracking architecture. I conduct multiple experiment on the framework, and provide detailed discussion of its viability and performance in different testing schemes. The source code and synthetic datasets used by this study can be found online.