Contour Completion Without Region Segmentation.

Area of Science:

• Computer Vision
• Computational Neuroscience
• Image Processing

Background:

• Contour completion is crucial for visual perception, grouping fragmented edges into coherent contours.
• Existing methods prioritize pixelwise accuracy, often neglecting the significant global contour closure effect.
• Psychological evidence highlights the importance of contour closure in human visual perception.

Purpose of the Study:

• To propose a novel, purely contour-based higher-order Conditional Random Field (CRF) model for contour closure.
• To address the limitations of existing methods by focusing on global contour closure rather than just pixelwise accuracy.
• To demonstrate that contour closure can be effectively achieved within the contour domain.

Main Methods:

• Development of a higher-order CRF model utilizing local connectedness approximation for contour closure.
• Transformation of the higher-order inference problem into an efficiently solvable integer linear program.
• Utilizing a bottom-up edge detector as a foundation for the proposed contour-based approach.

Main Results:

• The proposed method demonstrates superior contour grouping ability, as measured by the Rand index.
• Achieved comparable precision-recall performance relative to existing methods.
• Generated more visually pleasing results, highlighting effective contour closure.
• Successfully achieved contour closure directly in the contour domain.

Conclusions:

• Contour closure can be effectively achieved within the contour domain, without necessarily relying on image segmentation.
• The proposed higher-order CRF model offers an efficient and effective solution for contour completion and closure.
• This work challenges the conventional view that segmentation is a prerequisite for achieving contour closure.