Improving small object detection via cross-layer attention.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Small object detection is a critical challenge in computer vision.
• Feature pyramid networks (FPNs) are commonly used but have limitations in capturing long-range dependencies and handling feature noise.

Purpose of the Study:

• To propose a novel cross-layer attention (CLA) block to address limitations in FPN-based small object detection.
• To enhance the effectiveness of feature fusion by incorporating long-range interactions and reducing noise.

Main Methods:

• Developed a generic cross-layer attention (CLA) block for feature fusion.
• The CLA block considers both channel and spatial dimensions for reliable feature integration.
• Integrated the CLA block into existing state-of-the-art FPN-based object detection frameworks.

Main Results:

• The CLA block consistently improved performance in object detection and instance segmentation tasks.
• Experiments on the COCO 2017 dataset demonstrated the effectiveness of the proposed approach.
• The CLA block proved to be a lightweight and generalizable component.

Conclusions:

• The proposed cross-layer attention block effectively captures long-range dependencies and reduces noise in feature fusion.
• This method offers a significant advancement for small object detection and instance segmentation.
• The CLA block's versatility allows for easy integration into various feature fusion architectures.