PLS-GNN: 3D Water surface Object Detection Based on Point Cloud and Image fusion

Recently, two-stage detectors have shown strong application potential in the field of 3D object detection in water scenes. They usually use sampling keypoints or re-voxelization techniques to extract ROI features in the second stage. However, these methods often fail to fully exploit the rich information contained in the point cloud, especially the fine texture of the object surface and the deep correlation between points. In view of this limitation, this paper innovatively proposes PLS-GNN, a 3D object detection framework, aiming to achieve more accurate object detection from point cloud data. The PLS-GNN framework skillfully integrates voxel-wise neural networks with PointNet-based set abstraction techniques to capture and extract diverse features in point clouds. In addition, GNN-RoI pooling layer is introduced, which exploits the powerful ability of graph neural network to explicitly mine and integrate the texture information of object surface and the complex relationship between point clouds. Compared with traditional pooling methods, GNN-RoI can encode richer context information, which provides a solid foundation for accurately estimating the confidence and location of the target. Extensive experiments on the authoritative Waterscenes dataset fully verify the effectiveness of the framework. The experimental results show that PLS-GNN achieves significant improvement in detection accuracy, demonstrating its excellent ability to detect target objects efficiently and accurately in complex water scene 3D environment.