基于CenterNet的改进遥感旋转目标检测

刘鑫,黄进,杨瑛玮,李剑波

Improved Remote Sensing Rotating Object Detection based on CenterNet

Xin LIU,Jin HUANG,Yingwei YANG,Jianbo LI

表1 不同模型在HRSC2016数据测试集上精度评估结果

Table 1 Accuracy evaluation results of different models on HRSC2016 test data set

Models Backbone AnchorFree mAP@50（%） InputSize 两阶段算法 BL2[16] ResNet101 × 69.60 half of original size R2CNN[15] VGG16 × 73.07 800 x 800 RoI Tran[14] ResNet101 × 86.20 512 x 800 单阶段算法 IENet[35] ResNet101 √ 75.01 1024 x 1024 OriCenterness[36] ResNet101 √ 78.15 800 x 800 CSL[19] ResNet101 × 89.62 800 x 800 SAR[21] ResNet101 √ 88.45 896 x 896 RIE[22] HRGANet-W48 × 91.27 800 x 800 VR-CenterNet DLAs √ 88.48 640 x 640