Adaptive Feature Pyramid Networks for Object Detection

In general object detection, scale variation is always a big challenge. At present, feature pyramid networks are employed in numerous methods to alleviate the problems caused by large range of objects which makes use multi-level features extracted from backbone for top-down upsampling and fusion acquire set multi-scale depth image features. However, network proposed Ghiasi et al. adopts simple method, fails consider context, therefore, it difficult good addition, directly traditional prone misalignment loss details. this paper, an adaptive based on foregoing potential problems, includes two major designs, i.e., fusion. The aims predict group sampling points each pixel through some models, constitute representation combination points, while construct pixel-level weights between attention mechanism. experimental results verified effectiveness method paper. On public detection dataset MS-COCO test-dev, Faster R-CNN model achieved performance improvement 1.2 AP virtue network, FCOS could achieve 1.0 AP. What's more, experiments also validated that herein was more accurate localization.