Graphormer-Based Contextual Reasoning Network for Small Object Detection

Different from common objects, it is usually difficult to obtain fine enough visual information for small objects. Existing context-based learning methods only focus on whether the local contextual information of the object can be better extracted, but these methods easily lead to the loss of other effective information in the scene. To address this issue, this study proposes a Graphomer-based Contextual Reasoning Network (GCRN), which improves the feature expression of objects by modeling and reasoning about contextual information. Specifically, the method first establishes contextual information encoding using a context relation module, which is used to capture the local contextual features of objects as well as the dependencies among objects. Then, all the contextual information is aggregated by Graphormer in the context reasoning module to enrich the visual features of small objects. We conduct extensive experiments on two public datasets (i.e. MSCOCO and TinyPerson). The results show that this method can enhance the feature expression of small objects to a certain extent, and has good performance in feature information transmission.