Page - 116 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

synthesized versions of it. The authors utilize a SSIM based measure to adjust the reconstruction settings. This is close to the proposed approach in this paper. The main difference is that we do not need to process the available key frames but instead work only on synthetic views. III. AUTOMATIC PARAMETRIZATION The aim of automatic parametrization is to determine optimal values for different reconstruction parameter-types. In this case optimality means that the parameter value is near or equal the value a human operator would have chosen for the given data. Figure 1 depicts examples for the influences of different parameter-types. (a) ColourCorrection value 0. (b) ColourCorrection value 7. (c) Difference image. (d) PoseEstimation value 0. (e) PoseEstimation value 7. (f) Difference image. (g) SmoothnessTerm value 0. (h) SmoothnessTerm value 4. (i) Difference image. (j) MaxIterations value 0. (k) MaxIterations value 7. (l) Difference image. Fig. 1. Reconstruction effects of different parameter-types shown on Model 0001 in Subfigures (a) to (c), Model 0019 in Subfigures (d) to (f), Model 0010 in Subfigures (g) to (i), and Model 0033 in Subfigures (j) to (l). The images in the last column highlight the differences. The parameter values are set to the extremes, to better demonstrate the effects. Thefollowingprocedure illustrateshowanon-professional human operator could select a good parameter setting: 1) The operator sets or alters a parameter value. 2) The operator lets the reconstruction run. 3) The operator inspects the result from different views if it is better or worse than before. 4) The operator repeats the steps until some level of reconstruction quality is reached. Based on this we propose an approach using pair- wise view comparison of different reconstructions. For a parameter-typeα the accumulator matrixMα is a symmetric matrix defined as Mα(k,l) := n ∑ i:=0 SIM(Vi(Rα,k),Vi(Rα,l)) (4) where k,l are elements of the ordered parameter value set Pα andRα,k is the reconstruction. Elements inPα are chosen such there is an increasing influence of the parameter to the observed visual effect. n is the number of equally spaced views aroundRα,k, whereas each viewVi is a 2D projection of the 3D object. For comparison of two images as similarity SIM the Peak Signal to Noise Ratio (PSNR) is used. We assume that for humans the skin area is very important for quality judgement. Therefore the images are converted into the Hue, Saturation, Value (HSV) colour space before comparison.This shouldmake thecomparisonmoresensitive to skin parts, see Sedlacek [9]. A detailed evaluation and discussion of this step follows in Subsection IV-C and Section V. Pixel-wisecomparison isperformedonlyon thebust itself, since the background is masked out during comparison. Given this framework we propose the optimal parameter value oα∈Pα to be defined as oα :=argmax k∈Pα ∑ l∈Pα Mα(k,l). (5) Literally speaking the parameter value oα creates 2D views which are most similar to the views created with all other values. The hypothesis is that this is also a good parameter value which a human would choose. IV. EVALUATION Due to the lack of free datasets for bust reconstruction, an own dataset has been built up during an open house presentation in the company. A. Dataset The dataset contains 32 3D human bust scans showing different people, further called models. The data is acquired with a turntable and an off-the-shelf RGB-D sensor. Each individual is scanned in eight key poses. For the detailed set-up of the scan process see Heindl et al. [1]. For the reconstruction four different parameter-types are inspected: colourcorrection level, number of steps for pose estimation, surfacesmoothness termandmaximal iterationof the bundle adjustment. These types form the parameter set S={ ColourCorrection, PoseEstimation, SmoothnessTerm, MaxIterations}. Foradetailedexplanationof the reconstruc- tionsoftwareand theparameter semanticsseeagainHeindlet al. [1]. 116