Pushing the Boundaries of Boundary Detection using Deep Learning

In this work we show that Deep Convolutional Neural Networks can outperform humans on the task of boundary detection, as measured on the standard Berkeley Segmentation Dataset. Our detector is fully integrated in the popular Caffe framework and processes a 320x420 image in less than a second. Our contributions consist firstly in combining a careful design of the loss for boundary detection training, a multi-resolution architecture and training with external data to improve the detection accuracy of the current state of the art, from an optimal dataset scale F-measure of 0.780 to 0.808 while human performance is at 0.803. We further improve performance to 0.813 by combining deep learning with grouping, integrating the Normalized Cuts technique within a deep network. We also examine the potential of our boundary detector in conjunction with the higher level tasks of object proposal generation and semantic segmentation for both tasks our detector yields clear improvements over state-of-the-art systems.