Page - 115 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

On Quality Assurance of 3D Bust Reconstructions Gernot Stuebl, Christoph Heindl, Harald Bauer, and Andreas Pichler1 Abstract—In this paper a non-reference method for quality assurance in 3D bust reconstruction is presented. The proposed approach is part of an automatic parametrization concept for 3D reconstruction applications with no ground-truth data available. It is based on a novel concept of pair-wise view comparisons, which is new in this field. Evaluation on a dataset of human bust scans shows perfect prediction of human votes. I. INTRODUCTION Exact reconstruction of the human body especially the bust is an application field which got boosted by the raise of low-cost 3D printers and online 3D printing services. Nevertheless creating a high fidelity 3D reconstruction often involves manual post processing. Recent publications present systems which are able to do reconstructions on a quality level which makes post processing unnecessary, see Heindl et al. [1]. However for these the quality strongly relies on a correct parametrization of the system. Unfortunately parametrization is dependent on the scan data. So no golden standard for a parameter setting exists and the parameter values have to be adopted for each reconstruction individually. In principle human interaction has been shifted from direct manipulation/correction of 3D data to the selection of correct parameter values. Having this in mind, an (semi-)automatic configuration of the parameter values is desirable. The paper is outlined as followed: first Section II gives an overview of traditional quality assurance methods for 2D and follows with related work in the field of 3D quality assurance. The main approach is described in Section III, whereas Section IV presents the results on a dataset of 3D bust reconstructions. This is followed by a discussion on the applicability of the approach in Section V as well as a conclusion and outlook to future research in the last section. II. RELATED WORK A vital part of an automatic parametrization system is a component for assessing the reconstruction quality. The following subsections covers related work in this domain with an introduction of traditional 2D measures and the main emphasis on 3D quality assurance. A. 2DQuality Assurance In 2D there are traditional (dis-)similarity measures which are used for quality assurance. Some of these can also be 1PROFACTOR GmbH, 4407 Steyr-Gleink, Im Stadtgut A2, Austria {Forename.Surname}@profactor.at adopted to3D.Asimpleone is theRoot-Mean-SquaredError (RMSE) [5] of two images I,K which is defined as RMSE(I,K) := √√√√ 1 mn m−1 ∑ p=0 n−1 ∑ q=0 (I(p,q)−K(p,q))2 (1) and measures the deviation in each pixel. Based on this the Peak Signal to Noise Ratio (PSNR) [5] is defined as PSNR(I,K) :=20·log Imax RMSE(I,K) (2) with Imax the maximum possible value in the image (e.g. 255 for monochromatic 8bit images). PSNR measures the signal fidelity between an original and a disturbed image. A more complex measure is Structural Similarity index (SSIM) [2] which is designed to judge signal fidelity in the way the human vision system does. It is sensitive to structural distor- tions such as noise contamination, blurring, and insensitive to non-structural distortions such as luminance and contrast change. The mathematical definition is SSIM(~x,~y)= (2µxµy+c1)(2σxy+c2) (µ2x+µ2y+c1)(σ2x+σ2y+c2) (3) with c1=(k1L)2, c2=(k2L)2 as stabilization constants for the division with weak denominators, where L= 2b−1 denotes the dynamic range of pixel-values with b as the number of bits per pixel and k1=0.01 and k2=0.03. B. 3DQuality Assurance Generally, quality assurance algorithms are divided into full-reference (FR), reduced-reference (RR) and no-reference (NR) algorithms. This distinction is based on the amount of information that is available. Full-reference algorithms rely on a ground-truth data, e.g. early attempts to judge quality through texture and geometric resolutions belong to this category, see Pan et al.[3]. Also a broad range of algorithms which measure the quality of 3D codecs or stereoscopic 3D are full-reference based, see Mekuria et al. [4]. You et al. [5] give a good overview on how traditional 2D measures can be used for FR 3D quality assurance. For reduced-reference algorithms the ground-truth is not fully available. Instead of this, selected features are cal- culated from the ground-truth and used as input of the quality assurance system, see Wang et al. [6] or Rehman and Wang [7]. Arecent example forano-referencealgorithmispresented by Alexiadis et al. [8]. In this work the 2D key frames which are needed to build the 3D reconstruction are compared to 115