ClickAttention: Click region similarity guided interactive segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Interactive segmentation relies on user clicks, but current methods struggle with local focus and efficiency.
• Existing algorithms often require numerous clicks and face challenges in balancing performance and computational cost.

Purpose of the Study:

• To develop an efficient interactive segmentation algorithm that expands click influence and minimizes interference.
• To improve segmentation accuracy and reduce parameter count compared to state-of-the-art methods.

Main Methods:

• Proposed a click attention algorithm that leverages region similarity to broaden positive click impact.
• Introduced a discriminative affinity loss to decouple positive and negative click attention, preventing accuracy loss.
• Evaluated the method on the DAVIS dataset.

Main Results:

• Achieved a 2% performance gain (NoC@90) over SimpleClick-ViT-L on the DAVIS dataset.
• Reduced parameter usage to only 15.6% of the state-of-the-art method.
• Demonstrated superior performance and efficiency compared to existing interactive segmentation techniques.

Conclusions:

• The proposed click attention algorithm offers state-of-the-art performance in interactive segmentation with significantly fewer parameters.
• The method effectively addresses limitations of existing approaches, providing a more efficient and accurate solution.
• Published data and code facilitate further research and application in image segmentation.