Edge-aware joint neural denoising and normal estimation for mobile and handheld laser point clouds

Abstract. Portable laser scanners, handheld and mobile, have gained popularity for their ability to rapidly and economically document scenes. However, the acquired data are characterized by hight levels of noise and by low resolution, both affecting their consequent analysis and 3D modeling. It is customary to enhance their quality by denoising the data by means of point position updates where current approaches independently predict the displacement per point. Such strategies neglect local structural consistency and often yield a non-smooth outcome. To address these shortcomings this paper formulates denoising by local contextual relationships and point assignment to the underlying surface in an end-to-end framework. To extract contextual information it introduces densely packed graph convolution layers and a global attention mechanism. Realizing also that utilization of the conventional L2-norm-driven approaches tends to oversmooth the surface, it introduces a novel bilateral loss that not only mitigates the noise but also preserves sharp geometric features. As a result, our newly developed network learns a shape context representation that measures neighbor similarity and contributes to more accurate surface normals. Performance analysis demonstrates that over 93% of points deviate by ≤ 1 cm – double the percentage achieved by state-of-the-art denoising networks.