A path aggregation network with deformable convolution for visual object detection.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Object detection faces challenges with multi-scale objects.
• Existing methods for multi-scale feature fusion have limitations.

Purpose of the Study:

• To propose a novel network neck, DePAN, for effective multi-scale feature fusion in single-stage object detectors.
• To enhance the flexibility of feature point sampling using deformable convolutions.

Main Methods:

• Introduced the Deformable Convolution and Path Aggregation Network (DePAN).
• Integrated a deformable convolution block into the feature fusion branch of a path aggregation network.
• Implemented the deformable convolution block by stacking deformable convolution cells.
• Applied DePAN to Yolov6-N and YOLOV6-T baseline models.

Main Results:

• DePAN demonstrated effective fusion of multi-scale features.
• The proposed neck improved performance on COCO2017 and PASCAL VOC2012 datasets.
• Effectiveness was also validated on a medical image dataset.

Conclusions:

• The DePAN neck significantly enhances single-stage object detectors.
• DePAN offers flexibility and can be readily applied to various object detection models.
• The approach proves effective and applicable for real-world object detection tasks.