MilDetr: Detection Transformer for Military Camouflaged Target Detection

Abstract

Military Camouflage Target Detection (MCTD) is a special object detection task that aims to detect military camouflaged targets in the wild. In the challenging MCTD task, the confusing appearance and contours of military camouflaged targets often lead to the poor performance of existing methods. In this study, we propose an end-to-end Military Detection Transformer (MilDetr) for MCTD. We introduce two improvements to enhance the model’s performance. First, we employ the Reverse Features Feed Forward Neural Network (R3FN) for local information aggregation in the encoder of MilDetr. In addition, the Fusion Previous Query (FPQ) module is utilized for multi-stage query feature fusion in the decoder of MilDetr. To overcome data limitations for MCTD, we build two simulation military camouflaged target datasets called MilDet and MilCls. The ablation experiments on MilDet reveal the effectiveness of our improvements. Experimental results demonstrate that MilDetr obtains 95.6 AP on MilDet. Furthermore, MilDetr obtains 96.4 AP on MilDet with the pre-trained weights on ImageNet and MilCls. Compared with other object detectors, MilDetr achieves end-to-end military camouflaged target detection with superior performance.

Keywords

• Military camouflaged target
• object detection
• deep learning