Deep Crisp Boundaries: From Boundaries to Higher-Level Tasks.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep convolutional networks (ConvNets) have advanced edge detection, nearing human performance on benchmarks.
• Existing ConvNet edge detectors often produce imprecise edge localization, hindering downstream applications.
• Crisp edge maps are crucial for tasks requiring accurate boundary information.

Purpose of the Study:

• To systematically analyze the limitations of current ConvNet-based edge detectors.
• To develop a novel refinement architecture for learning accurate, crisp edge detectors.
• To demonstrate the practical benefits of improved edge maps in computer vision.

Main Methods:

• A systematic study of ConvNet-based edge detector outputs was conducted.
• A novel top-down backward refinement architecture was proposed.
• The architecture progressively increases feature map resolution to generate crisp edges.

Main Results:

• The proposed method generates significantly crisper edge maps compared to existing approaches.
• Performance surpasses human accuracy on the BSDS500 benchmark using standard criteria.
• Outperforms state-of-the-art methods under stricter evaluation criteria.

Conclusions:

• The novel refinement architecture effectively learns crisp edge detectors using ConvNets.
• Crisp edge maps derived from this method offer substantial improvements for computer vision tasks.
• Applications like optical flow estimation, object proposal generation, and semantic segmentation benefit from precise edge localization.