End-to-end object detection method based on shape correction and feature selection cross-attention

DEtection TRansformer (DETR) and its variant models degrade the object detection performance due to the inability to provide the object position a priori and the lack of shape deviation supervision between prediction boxes and ground truth boxes, which makes some of the shape differences between prediction boxes and ground truth boxes too large. To address this problem, we propose an end-to-end object detection method (SF-DETR) based on shape correction and feature selection cross-attention. First, a feature selection layer is introduced to provide the model with a priori object position by combining classification confidence and intersection over union score to filter high-quality prediction boxes and constrain the computational range of decoder cross-attention. Second, the shape bias is numericalized and a shape correction loss function is proposed to supervise the shape bias and ensure that the shape bias is effectively corrected. Finally, we conducted experiments on three public datasets, MS COCO, Caltech Pedestrians, and BIT-Vehicle, and the experimental results show that the method in this paper not only provides the object position a priori effectively but also reduces the shape bias significantly, which improves the prediction box quality and object detection accuracy.