3D Object Detection based on Surround-View Fisheye Cameras

Fisheye cameras, due to their wide field of view, have been widely applied in the advanced driver assistance systems of commercially available vehicles. However, 3D object detection research based on fisheye images remains relatively scarce. There are two primary factors: firstly, existing fisheye image datasets lack precise annotations for 3D objects, thereby rendering them unsuitable for supervised training; secondly, the severe distortion of fisheye images poses challenges in learning geometry mapping for 3D object detection model. In our work, we establish an annotated surround-view fisheye dataset specifically for 3D object detection. We also propose the corrected fisheye distortion module and build the Mono-Fisheye, a wide-angle 3D object detection model based on monocular fisheye images. Additionally, we propose a BEV 3D object detection model based on surround-view fisheye images named Fisheye-BEVDet, and validate the effectiveness of related data augmentation strategies.