PROJECT TITLE :

Deep Crisp Boundaries From Boundaries to Higher-Level Tasks

ABSTRACT:

It has become possible thanks to the use of deep convolutional networks to improve edge detection (ConvNet). The performance of these ConvNet-based edge detectors on conventional benchmarks has come close to that of a human. These detectors' outputs are studied in detail. If a task requires precise edge input, we show that the detection results did not precisely locate edge pixels. Refinement architecture is proposed to address the tough challenge of learning a crisp detector with ConvNets. As the resolution of feature maps is gradually increased, the edges of the images get sharper and sharper. When utilising normal criteria on BSDS500, our results surpass human accuracy, and when employing stricter criteria, they largely outperform state-of-the-art approaches. Furthermore, we demonstrate the advantages of sharp edge maps for a variety of computer vision applications, including optical flow prediction, object proposal generation and semantic segmentation.