Dynamic Feature Integration for Simultaneous Detection of Salient Object, Edge and Skeleton

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing methods for salient object segmentation, edge detection, and skeleton extraction are task-specific, leading to inefficiencies in storage and sequential processing.
• Jointly solving related tasks often requires datasets with multiple, simultaneously supported annotation types, which are not always available.

Purpose of the Study:

• To develop a unified, end-to-end trainable framework for multiple low-level pixel-wise vision tasks.
• To demonstrate the shared characteristics of salient object segmentation, edge detection, and skeleton extraction and leverage them for a single model.

Main Methods:

• Introduced a selective integration module enabling dynamic feature selection from a shared backbone based on task-specific requirements.
• Designed a task-adaptive attention module for intelligent information allocation tailored to image content priors.
• Developed a unified network trained end-to-end for multiple vision tasks.

Main Results:

• The proposed network achieved strong performance across salient object segmentation, edge detection, and skeleton extraction tasks.
• Outperformed current single-purpose state-of-the-art methods on multiple representative datasets.
• Ablation analyses validated the design principles of the unified framework.

Conclusions:

• A unified framework can effectively address distinct pixel-wise vision tasks like segmentation, edge detection, and skeleton extraction.
• The proposed selective integration and task-adaptive attention modules are key to the network's versatility and performance.
• The research facilitates future work by providing a versatile model and releasing source code.