Fast Edge Detection Using Structured Forests.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Edge detection is fundamental for computer vision tasks like object detection and image segmentation.
• Local image patches contain structural information (lines, junctions) that can be leveraged for improved edge detection.

Purpose of the Study:

• To develop an accurate and computationally efficient edge detector by exploiting local image patch structure.
• To formulate edge mask prediction within a structured learning framework using random decision forests.

Main Methods:

• Utilized a structured learning framework applied to random decision forests for edge detection.
• Developed a novel method for learning decision trees that maps structured labels to a discrete space for information gain evaluation.

Main Results:

• Achieved real-time performance, significantly faster than existing state-of-the-art methods.
• Obtained state-of-the-art edge detection results on the BSDS500 Segmentation and NYU Depth datasets.
• Demonstrated strong generalization capabilities of learned edge models across different datasets.

Conclusions:

• The proposed structured learning approach offers a computationally efficient and accurate solution for edge detection.
• The method's ability to generalize makes it a versatile tool for various computer vision applications.