Adaptive Linear Span Network for Object Skeleton Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Conventional object skeleton detection networks often rely on manual design, lacking theoretical grounding and requiring extensive expertise.
• Existing methods struggle with integrating object/part representations across different granularities.

Purpose of the Study:

• To propose an automated method for object skeleton detection using neural architecture search (NAS).
• To develop an Adaptive Linear Span Network (AdaLSN) that integrates scale-aware features and utilizes the theory of linear span for multi-scale feature fusion.

Main Methods:

• Formulated AdaLSN based on the theory of linear span for multi-scale deep feature fusion.
• Defined a mixed unit-pyramid search space, enabling joint optimization of unit-level operations and pyramid-level connections.
• Applied genetic architecture search to adaptively expand feature spaces.

Main Results:

• Achieved a superior accuracy and latency trade-off compared to state-of-the-art methods.
• Demonstrated AdaLSN's versatility and effectiveness in general image-to-mask tasks like edge detection and road extraction.

Conclusions:

• AdaLSN offers an automated, theoretically grounded approach to object skeleton detection.
• The proposed method provides significant improvements in performance and adaptability for various computer vision tasks.