A dual-path network for semantic scene completion of single-frame LiDAR point clouds

Semantic Scene Completion (SSC) is a fundamental yet challenging task in 3D environment perception, as the sparsity and noise of LiDAR point clouds make it difficult to accurately recover both geometry and semantics. To address these challenges, we propose DPS2CNet, a novel Dual-Path SSC Network that integrates voxel-based and bird’s-eye view (BEV) representations to exploit their complementary strengths. Specifically, DPS2CNet employs a Cylinder3D-enhanced voxel branch to capture fine-grained 3D geometry and a UNet-based BEV branch to model large-scale contextual information. To further boost performance, we incorporate CARAFE for efficient feature upsampling and design a tailored loss function optimized for SSC. Extensive experiments on SemanticKITTI and SSCBench-KITTI-360 demonstrate that DPS2CNet achieves state-of-the-art results. In particular, it ranks first on the SemanticKITTI test set with an IoU of 62.6% among all open-source submissions 1 , highlighting its effectiveness in complex real-world driving scenarios.