AIoU: Adaptive bounding box regression for accurate oriented object detection

Abstract

Object detection consists of two key steps: class recognition and object localization. Class recognition is fundamental because the quality of the obtained feature representations is key to detection accuracy; for locating the object, bounding box refinement is the most intuitive method for improving the localization accuracy of the utilized detector; that is, selecting a better loss function metric when computing the best‐fitted bounding box for the object of interest. However, current class activation mapping (CAM) scores cannot help effectively distinguish the object from background noise and involve fixed weights for the geometric characteristics of the anchor box, leading to inaccurate object detection. In this paper, we proposed a mixed‐CAM method to obtain improved category scores for class recognition, and an adaptive intersection‐over‐union method (AIoU) that improves the localization performance for object detection. The mixed‐CAM method combines an original image response and CAM information to provide a confidence score for the final feature map and, in the meantime, considers this score as a sample selection criterion for the following localization regression stage. The AIoU method designs a new loss function metric for bounding box localization regression. In doing so, the proposed method considers the weight of each geometric characteristic of the bounding box in the network training process via a hyperparameter and adopts a new positive and negative sample selection mechanism for sample training. Experimental results show that the proposed framework achieves better prediction accuracy and a higher average precision value than those yielded by the classical backbone networks. Moreover, the AIoU method can be easily coupled with existing convolutional neural network architectures and thus possesses the great potential of adaptability in many application fields, such as intelligent transportation.