Improved Dual Attention for Anchor-Free Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Anchor-free object detection methods often prioritize bounding box centers.
• The significance of central regions can vary, impacting detection accuracy.
• Existing attention mechanisms may confuse spatial and channel information.

Purpose of the Study:

• To propose a novel dual attention-based approach for adaptive weight assignment in anchor-free object detection.
• To untie spatial and channel attention for improved weight determination.
• To enhance the overall correctness and accuracy of object detection.

Main Methods:

• Developed an end-to-end network including backbone, feature pyramid, adaptive weight assignment, regression, and classification.
• Implemented a parallel framework for dual attention using depthwise convolution for spatial attention and 1D convolution for channel attention.
• Utilized depthwise convolution to prevent interference between spatial and channel attention, and 1D convolution for efficiency and effectiveness.

Main Results:

• Achieved an average precision of 52.7% on the MS-COCO dataset.
• Demonstrated significant improvement compared to other anchor-free object detectors.
• The proposed dual attention mechanism effectively assigns proper attention weights.

Conclusions:

• The novel dual attention mechanism enhances object detection accuracy by adaptively weighting bounding box regions.
• Separating spatial and channel attention through specific convolutional approaches improves performance.
• The method offers a significant advancement in anchor-free object detection capabilities.