While the duality between boundary and medial representations has been exploited in the context of pre-segmented shapes, it has not been studied in the context of natural scenes. The goal of this project is to use a shared feature representation to address edge and skeleton detection simultaneously, improving performance for both tasks. To compare the relative benefits of a joint approach, we plan to use a single convolutional neural network (CNN) for both tasks, combined with a novel loss function that enforces consistency between detected boundary and medial points.

Outgoing