OrtDet: An Orientation Robust Detector via Transformer for Object Detection in Aerial Images

The detection of arbitrarily rotated objects in aerial images is challenging due to the highly complex backgrounds and the multiple angles of objects. Existing detectors are not robust relative to the varying angle of objects because the CNNs do not explicitly model the orientation's variation. In this paper, we propose an Orientation Robust Detector (OrtDet) to solve this problem, which aims to learn features that change accordingly with the object's rotation (i.e., rotation-equivariant features). Specifically, we introduce a vision transformer as the backbone to capture its remote contextual associations via the degree of feature similarities. By capturing the features of each part of the object and their relative spatial distribution, OrtDet can learn features that have a complete response to any direction of the object. In addition, we use the tokens concatenation layer (TCL) strategy, which generates a pyramidal feature hierarchy for addressing vastly different scales of objects. To avoid the confusion of angle regression, we predict the relative gliding offsets of the vertices in each corresponding side of the horizontal bounding boxes (HBBs) to represent the oriented bounding boxes (OBBs). To intuitively reflect the robustness of the detector, a new metric, the mean rotation precision (mRP), is proposed to quantitatively measure the model's learning ability for a rotation-equivariant feature. Experiments on the DOTA-v1.0, DOTA-v1.5, and HRSC2016 datasets show that our method improves the mAP by 0.5, 1.1, and 2.2 and reduces mRP detection fluctuations by 0.74, 0.56, and 0.52, respectively.