6-DoF Pose Estimation of Marine Objects Using Geometrically Constrained 3D Bounding Boxes

Six-degree-of-freedom (6-DoF) pose estimation of marine targets is essential for maritime and underwater operations. To address the instability caused by complex monocular marine scenes, this paper proposes GeoBox-Pose, a 6-DoF pose estimation method based on geometrically constrained 3D bounding boxes. The method constructs a 3D bounding-box representation from target physical dimensions, reducing dependence on accurate 3D models. A shared visual encoder with a multi-branch regression network is designed to jointly predict 2D projected corners, 3D orientation, and spatial translation. A projection-consistency constraint based on the camera imaging model is further introduced to reproject the pose-driven 3D bounding box onto the image plane, enhancing consistency between image features and 3D geometry. The results show that GeoBox-Pose achieves mean rotation errors of 14.96° and 12.36° on the BONK-pose and ROV6D datasets, respectively, with corresponding mean projection errors of 100.84 px and 75.63 px. The RotAcc@20° values reach 85.74% and 91.76%, while the ProjAcc@100px values reach 81.95% and 84.05%, respectively. The model contains 11.67 M parameters and achieves an inference speed of nearly 300 FPS. The experiments further confirm that the projection-consistency constraint contributes to improved pose-estimation performance. Overall, GeoBox-Pose enables monocular 6-DoF pose estimation for different types of marine targets using a unified 3D bounding-box geometric representation. While maintaining favorable pose-estimation performance, it substantially reduces model size and computational cost and achieves high network inference efficiency, thereby providing a good overall balance between pose-estimation accuracy and network inference efficiency. The proposed method offers a lightweight geometry-constrained pose-estimation solution for marine engineering tasks such as ship attitude monitoring and underwater-robot localization.