A path aggregation network with deformable convolution for visual object detection

One of the main challenges encountered in visual object detection is the multi-scale issue. Many approaches have been proposed to tackle this issue. In this article, we propose a novel neck that can perform effective fusion of multi-scale features for a single-stage object detector. This neck, named the deformable convolution and path aggregation network (DePAN), is an integration of a path aggregation network with a deformable convolution block added to the feature fusion branch to improve the flexibility of feature point sampling. The deformable convolution block is implemented by repeated stacking of a deformable convolution cell. The DePAN neck can be plugged in and easily applied to various models for object detection. We apply the proposed neck to the baseline models of Yolov6-N and YOLOV6-T, and test the improved models on COCO2017 and PASCAL VOC2012 datasets, as well as a medical image dataset. The experimental results verify the effectiveness and applicability in real-world object detection.