Delving Into Crispness: Guided Label Refinement for Crisp Edge Detection.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Learning-based edge detection models often produce thick edges.
• Existing methods struggle with edge crispness due to label noise.

Purpose of the Study:

• To investigate the cause of thick edge predictions in learning-based edge detection.
• To propose a method for improving edge crispness by refining training labels.
• To demonstrate the effectiveness of refined labels for training crisp edge detectors.

Main Methods:

• Developed a novel edge crispness measure for quantitative evaluation.
• Proposed a Canny-guided refinement technique for human-labeled edges.
• Trained existing edge detection models using the refined edge maps.

Main Results:

• Identified noisy human labels as the primary cause of thick edge predictions.
• Refined edge maps significantly improved edge crispness in trained models (17.4% to 30.6%).
• Achieved state-of-the-art performance on the Multicue dataset with improved ODS (12.2%) and OIS (12.6%) using the PiDiNet backbone without non-maximal suppression.

Conclusions:

• Prioritizing label quality over model design is crucial for achieving crisp edge detection.
• The proposed Canny-guided edge refinement effectively enhances training data for crisp edge detectors.
• Crisp edge detection demonstrates superior performance in downstream applications like optical flow estimation and image segmentation.