DGTPNet: Enhancing point cloud completion with 3D Gaussian representations under simulated occlusion and fog

Point cloud shape completion is a challenging task essential for numerous 3D applications, including robotics and autonomous driving. Existing datasets often lack realism, leading models to specialize in idealized missing patterns. To address this, we introduce a data simulation method that incorporates structural occlusions and fog-induced sparsification and noise into the ShapeNet-55 dataset, better aligning it with LIDAR-generated data. We also propose DGTPNet, an encoder–decoder architecture that employs 3D Gaussian distributions as an intermediate representation. Rather than regressing point coordinates directly, DGTPNet reconstructs the complete object geometry using Gaussian primitives, from which the final point cloud is sampled. Experimental results demonstrate that our model achieves a CD-L2 (Avg) of 1.03 on the standard ”Online Cropping” benchmark and 1.794 on our proposed ”Occlusion and Fog” benchmark. Notably, in the ”Heavy” fog setting, our model maintains a CD-L2 of 2.194, indicating strong robustness and stable reconstruction performance for point cloud completion in highly challenging scenarios.