Boundary-Aware Supervoxel-Level Iteratively Refined Interactive 3D Image Segmentation With Multi-Agent Reinforcement...

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing interactive segmentation methods often overlook the sequential nature of user interactions.
• Treating each interaction independently limits the efficiency and accuracy of segmentation mask generation.

Purpose of the Study:

• To develop an advanced interactive image segmentation method using reinforcement learning.
• To improve segmentation accuracy and efficiency by modeling the dynamics of successive user interactions.

Main Methods:

• Modeling iterative interactive image segmentation as a Markov decision process (MDP).
• Employing multi-agent reinforcement learning (MARL) with a shared voxel-level policy.
• Introducing a boundary-aware reward incorporating global cross-entropy gain and boundary prediction weights.
• Utilizing a supervoxel-clicking interaction design combining point-clicking and scribbles.

Main Results:

• The proposed method significantly outperforms state-of-the-art techniques on benchmark datasets.
• Achieved higher accuracy in generating segmentation masks.
• Demonstrated enhanced robustness and required fewer user interactions.

Conclusions:

• The MARL framework effectively captures the dynamics of iterative user interactions for image segmentation.
• The boundary-aware reward mechanism improves the precision of segmentation, particularly at object boundaries.
• The supervoxel-clicking interaction design offers a balanced approach to interaction efficiency and robustness.