EFR-FCOS: enhancing feature reuse for anchor-free object detector

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Object detection is crucial for computer vision tasks.
• Fully convolutional one-stage object detection models face challenges in feature reuse.
• Enhancing feature extraction and fusion is key to improving detection accuracy.

Purpose of the Study:

• To propose an enhanced feature reuse method (EFR-FCOS) for fully convolutional one-stage object detection.
• To improve feature extraction in the backbone, feature fusion in the neck, and detection in the head.
• To achieve significant performance gains in object detection.

Main Methods:

• Global Attention Network (GANet) for prominent feature extraction in the backbone.
• Aggregate Feature Fusion Pyramid Network (AFF-FPN) with attention for improved feature fusion in the neck.
• Cascade detection with refined bounding box regression in the head (EnHead) for enhanced classification and regression.

Main Results:

• The proposed EFR-FCOS method demonstrates extensive usability.
• Significant performance improvements were achieved on the COCO object detection dataset.
• GANet, AFF-FPN, and EnHead collectively enhance object detection capabilities.

Conclusions:

• The EFR-FCOS framework effectively enhances feature reuse across object detection components.
• The integration of global attention and feature fusion techniques leads to superior detection performance.
• The proposed methods offer a promising direction for advancing one-stage object detection models.