Holistically-Attracted Wireframe Parsing: From Supervised to Self-Supervised Learning.

Area of Science:

• Computer Vision
• Geometric Analysis
• Machine Learning

Background:

• Wireframe parsing is crucial for understanding 2D image structure.
• Existing methods often struggle with accuracy and efficiency.

Purpose of the Study:

• To introduce Holistically-Attracted Wireframe Parsing (HAWP) for robust 2D wireframe analysis.
• To enhance wireframe parsing performance in both supervised and self-supervised settings.

Main Methods:

• HAWP employs a Holistic Attraction (HAT) field representation for line segments.
• It uses a three-component pipeline: HAT field to line segment generation, segment-to-endpoint binding, and an endpoint-decoupled line-of-interest aligning (EPD LOIAlign) module for verification.

Main Results:

• HAWPv2 demonstrates strong performance in fully supervised learning.
• HAWPv3 achieves superior repeatability scores and efficient training in self-supervised learning.
• HAWPv3 shows potential for out-of-distribution wireframe parsing without ground truth labels.

Conclusions:

• HAWP provides an effective and efficient approach to wireframe parsing.
• The HAT field representation and novel modules significantly improve performance.
• HAWP, particularly in its self-supervised variant, offers a promising direction for future research in geometric analysis.