Mesh Denoising via Cascaded Normal Regression

摘要

We present a data-driven approach for mesh denoising. Our keyrnidea is to formulate the denoising process with cascaded non-linearrnregression functions and learn them from a set of noisy meshes andrntheir ground-truth counterparts. Each regression function infers thernnormal of a denoised output mesh facet from geometry featuresrnextracted from its neighborhood facets on the input mesh and sendsrnthe result as the input of the next regression function. Specifically,rnwe develop a filtered facet normal descriptor (FND) for modeling therngeometry features around each facet on the noisy mesh and modelrna regression function with neural networks for mapping the FNDsrnto the facet normals of the denoised mesh. To handle meshes withrndifferent geometry features and reduce the training difficulty, werncluster the input mesh facets according to their FNDs and train neuralrnnetworks for each cluster separately in an offline learning stage.rnAt runtime, our method applies the learned cascaded regressionrnfunctions to a noisy input mesh and reconstructs the denoised meshrnfrom the output facet normals.rnOur method learns the non-linear denoising process from the trainingrndata and makes no specific assumptions about the noise distributionrnand geometry features in the input. The runtime denoising processrnis fully automatic for different input meshes. Our method can berneasily adapted to meshes with arbitrary noise patterns by training arndedicated regression scheme with mesh data and the particular noisernpattern. We evaluate our method on meshes with both syntheticrnand real scanned noise, and compare it to other mesh denoisingrnalgorithms. Results demonstrate that our method outperforms thernstate-of-the-art mesh denoising methods and successfully removesrndifferent kinds of noise for meshes with various geometry features.