PVLF: point-voxel local feature fusion for 3D detection

Abstract Significant progress on 3D object detection in point cloud has been made in the detection of large objects with clear shape and contour information, such as cars. However, existing algorithms still face significant challenges in detecting tiny objects such as cyclists and pedestrians. This paper presents a novel feature fusion method named Point Voxel Local Feature Fusion (PVLF), which deeply integrates point cloud and voxel information. PVLF explores local spatial features to improve accuracy regarding tiny object detection. To address the potential complex computational issues in the convolution process, we have designed an innovative Adaptive Sparse Convolution (ASC) module that effectively eliminates redundant information in the feature layer. Due to long-range dependencies in point cloud feature extraction, the Dynamic Graph Convolution combined with Transformer (DGFormer) is developed as the point cloud feature encoder. DFFormer expands the receptive field to capture contextual information and could hence improve deep representation learning. We also introduce a sector segmentation based sampling strategy and achieves parallel sampling of key points through our designed Adjacency Distance Update Farthest Point Sampling (ADUFPS) algorithm, significantly reducing computational overhead while improve sampling efficiency. The experimental results on the KITTI and Waymo datasets show that our method outperforms, particularly in tiny object detection tasks, the-state-of-the-arts deep models which are also based on the point-voxel feature fusion.