A Dual Relation Extractor for Object Detection

It is well known that context can help object detection, but mainstream single-stage object detection algorithms still detect object instances individually. In this work, we propose a Relation Extraction Module (REM) which extracts both global context and local context at the same time. It processes a set of anchors simultaneously through interaction between their appearance and spatial features, thus allowing building local context. It gets the global context by filtering and averaging all anchor features of the current feature layer, thus allowing building background information of the image. It does not need additional manual labeling information and is easy to plug into popular detectors. Experiments on MS COCO datasets indicate that REM improves the accuracy of the popular single-stage detector while maintains real-time performance.