TRANSFORMER-BASED ARBITRARY ORIENTED OBJECT DETECTION IN VERY-HIGH-RESOLUTION REMOTE SENSING IMAGES

In remote sensing (RS) field, arbitrary oriented object detection (AOOD) technology is dominated with convolutional neural networks (CNNs). However, CNN-based AOOD technologies suffer from the structural limitations for capturing the contextual relationships in RS image. To address this issue, we propose a transformer-based AOOD method capable of effectively capturing and understanding the rotational characteristics of individual objects. Specifically, we replaced certain convolutional layers in the CNN backbone with a module capable of dynamically generating feature maps based on the directional characteristics of objects. Additionally, we adopted a detection transformer (DETR) framework as the detector, enabling the generation of oriented object queries and designing a network that effectively considers both local and global rotational characteristics of objects. To demonstrate the effectiveness of the proposed approach, we trained the network on an open-source dataset including very-high-resolution (VHR) images for AOOD tasks. The results showed that the proposed method achieved superior region proposal accuracy and rotation angle estimation performance compared to other networks.