Selected and refined local attention module for object detection

Lacking enough feature expression on the shallow part of the network always hinders the object detection results by missing the position of small instances. To address this, the authors propose a selected and refined local attention module (SRLAM) for object detection. SRLAM tries to improve the feature expression of local areas by establishing a relationship graph between different channels. Inspired by the classical non-local neural networks for video classification, they present the local attention module (LAM) to more effectively use remote information. The LAM can suppress the influence of irrelevant areas for improving detection accuracy. Moreover, for further information utilisation, the weight map generated by LAM is integrated together with the original feature map and the information from the higher layer by a feature fused module. Experiments on the PASCAL VOC2007 show that the authors' model has good detection performance, the proposed model can achieve 81.6 mAP when the size of the input image is 300 x 300, that outperforms many other mainstream detectors.