Feature Fusion and Information Supervision Consistency for Object Detection

The inconsistency between classification and regression is a common problem in the field of object detection. Such may lead to undetected objects, false detection, boxes overlapping detection results. It has been determined that mainly caused by feature coupling lack information regarding interactions heads. In this study, characteristics spatial invariance were used, ability fit data distribution was enhanced fully connected layers. A fusion module (FFM) proposed order enhance capabilities model’s extractions. This study also further considered loss functions function (RMAE) based on mean absolute error (MAE) for purpose improving location quality. Furthermore, solve heads, an (Lin) added basis module. Then, evaluate effectiveness methods, network (FMRNet) trained RetinaNet. experimental results demonstrated study’s methods surpassed accuracy some existing detectors when FMRNet adopted. confirmed had problems overlapping.