On Object Detection Method of Tank Base on SNMS
Xiaozhu Xie, Hai-rong Hu, Yang Zhang, Cheng He · 2018 IEEE CSAA Guidance, Navigation and Control Conference (CGNCC) · 2018
The detection of overlapping tank targets in visible battlefield reconnaissance image was mainly studied. Analyzed the problem of the target detection of the Faster RCNN based on NMS and proposed the method of SNMS to improve the precision of detection and positioning of tank target detection in overlapping cases. In the process of bounding box regression, the traditional non-maximum suppression[4]is difficult to detect the overlapping tank object by hard threshold decision. We used the soft threshold method based on the score-resetting function to preserve the overlapping region proposals as many as possible. The experimental results show that the proposed method can significantly improve the performance of the overlapping tank object without increasing the computational complexity and additional training.