Abstract: To address arbitrary-oriented object detection in remote sensing imagery,a cascaded object detec-tion algorithm is proposed.Based on the baseline model,this model adopts two bounding box labeling modes to mutually correlate multiple regions of interest networks.Then,with the prediction results of the oriented bounding box of the current stage,it predicts the horizontal bounding box and the oriented bound-ing box in the next stage,forming a multi-stage cascaded learning process.In addition,this model combined the advantages of the horizontal bounding box mode and the oriented bounding box mode to achieve more robust predictions.A large number of experimental results on the DOTA dataset show that the bounding box prediction accuracy of the proposed algorithm is superior to the existing detection algorithms for remote sensing imagery on both two bounding box tasks.