TransNet: Linkage Recognition with Contextual Nexus for Hard Object

摘要

One of the major challenges in object recognition is to propose a structure with highly performance of overlapping, small and misclassification objects, called hard object recognition. There is a contextual nexus among the spatial location and geometric characteristics of objects, which we find is essential to the recognition efficacy. We propose regress connection, nexus calculation module and linkage loss, which are contained in an improving design TransNet framework. It can not only demonstrates the interactivity of explicit nexus between objects, but also achieve better results through the integration of two mainstream methods. Adjusting the training attention through linkage loss function, thus significantly enhance the performance of hard objects. The experiment results show that our method achieves remarkable performance on PASCAL VOC2012 and MS COCO data set. Also, the head network is capable of integrating with region-based methods and gains better performance.