FGF3D: A Point Cloud 3-D Object Detection Algorithm Based on Fine-Grained Feature Fusion

In the existing two-stage point cloud 3D object detection task, there are problems such as the lack of fine-grained and insufficient sparsity of point cloud features extracted by the backbone network, the lack of spatial-channel fine-grained fusion of bird-eye-view (BEV) features, and the lack of multi-dimensional representation of ROI pooling features. We innovatively propose a new two-stage algorithm based on fine-grained feature fusion, called FGF3D. We construct a sparsegrained aggregation network (SGANet) to elevate the representation of point cloud features. Furthermore, we innovatively propose a spatial-channel fine-grained fusion module (SCFFM) to enrich BEV features and efficiently extract fine-grained features. Further, we propose a global-local feature pooling fusion module (GLFP), which effectively fuses local detail object features and global features, enhances multi-dimensional representations, and facilitates the refined regression of proposal boxes. Without bells and whistles, FGF3D attains excellent performance on two autonomous driving benchmarks, with detection accuracies of up to 64.10% on KITTI, 57.82% on ONCE.