Attention Based Part-Aware Features of Anchor Boxes for Single-Shot Object Detection

In this paper,we propose a lightweight but effective design to enhance the anchor representations for improving the performance of single-shot object detectors.This design consists of an attention module and a part-aware module.First,an attention module is applied for a location to adaptively express various representations based on the targets of the anchor boxes covering the location.The part-aware module further extracts the discriminative part features inside each anchor box as its individual features for robust prediction.Assembling the proposed modules to the SSD,our method is able to consistently boost the performance on several public benchmarks while maintaining real-time inference speed(14 ms).Extensive experiments on the region proposal generation indicate that the proposed method also promotes the recall of region proposals,so as the accuracy of two-stage object detection.