The objects in remote sensing images are usually small and dense with complex background. The most important feature is non-axial symmetry and arbitrary alignment. Therefore, HBB (horizontal bounding box) is not suitable to represent the objects in remote sensing images. We propose a detection framework named Oriented CenterNet, which can detect arbitrary orientation objects efficiently. We adopt the light weight backbone called Swin-Tiny Transformer. Compared with CNN-based backbone, it can obtain the global receptive field and establish connections with other pixel points. And then, we propose a new and simple six-parameter representation, named Cornerpoint Offset Representation, to represent rotated objects, which can convert HBB to RBB (rotated bounding box) easily. For small objects in remote sensing images, the network adopts an effective feature fusion and sampling method, Center Pooling is integrated into the prediction module, to enhances the features of small objects and weakens the background noise. In addition, we redefine the loss function by dividing the regression loss into horizontal and rotated parts, in order to provide better supervision in the training process of the network. Experiments illustrate the effectiveness of our method. Our Oriented CenterNet with Swin-Tiny-FPN achieving 70.21 % mAP on DOTA dataset.
Paper
Full text
Oriented CenterNet: Rotated Object Detection in Remote Sensing Images
Semantic Scholar · Environmental Science · 2023
Abstract
The objects in remote sensing images are usually small and dense with complex background. The most important feature is non-axial symmetry and arbitrary alignment. Therefore, HBB (horizontal bounding box) is not suitable to represent the objects in remote sensing images. We propose a detection framework named Oriented CenterNet, which can detect arbitrary orientation objects efficiently. We adopt the light weight backbone called Swin-Tiny Transformer. Compared with CNN-based backbone, it can obtain the global receptive field and establish connections with other pixel points. And then, we propose a new and simple six-parameter representation, named Cornerpoint Offset Representation, to represent rotated objects, which can convert HBB to RBB (rotated bounding box) easily. For small objects in remote sensing images, the network adopts an effective feature fusion and sampling method, Center Pooling is integrated into the prediction module, to enhances the features of small objects and weakens the background noise. In addition, we redefine the loss function by dividing the regression loss into horizontal and rotated parts, in order to provide better supervision in the training process of the network. Experiments illustrate the effectiveness of our method. Our Oriented CenterNet with Swin-Tiny-FPN achieving 70.21 % mAP on DOTA dataset.