Seite - 154 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Fig. 3: Schematic illustration of the Blender Node Setup for creating different materials. photometric stereo with five light sources in order to recover the3Dsurfaceofanobject.Usingepipolarplane image(EPI) structures from motion analysis for depth reconstruction was introduced in [9]. The paper by Tao deals with incorporating a shading term to depth from defocus with correspondence cues in order to refine the shape estimation [10]. In [11] Hayakawa used a singular-value decomposition (SVD) of a formulated matrix in order to get a surface normal estimation without the need of a-priori knowledge of the light source direction under the Lambertian assumption. Some machine learning approaches have been explored, such as [12] where a multi-layered neural network was used in order to learn the mapping between image intensities and the surface normal orientation, using a Gaussian sphere with average reflectance as the training data. In [13] Cheng used a symmetrical 6-layer neural network to train a mapping between the vectorized image and a reflectance value for each pixel. Another machine learning approach has been investigated in [14],whereaneuralnetworkwasused inorder tosolve the shape from shading problem, previously introduced by [15]. III. EXPERIMENTS A. GeneratingGround Truth Data Blender 2.78 [16] Cycles Renderer was used to generate the ground truth data. This artificial ground truth data has some advantages over real-world acquisition, such as the ease of modification of the setup, feasibility of generating many images quickly as well as being less prone to errors. However, in order to make the resulting images more re- alistic, some artefacts, such as jitter or salt&pepper noise, can be taken into consideration. The goal while creating the ground truth data was to cover as much ground as possible with the synthetic data regarding the task. The network should learn a mapping between the RGB intensity vectors of the different views and surface properties, to the surface normal gradient. As it is infeasible to cover all possible map- pings between color, light reflectance and surface normals, Fig. 4: Visualization of the six different datasets created by changing the roughness and the percentage of which the glossy or diffuse node is taken. a random approach was chosen. A uniformly distributed, 8-bit random color pattern was created (each RGB color channel uniformly distributed between 0-255) and used as a texture. The blender-intern noise texture and displacement map node was used in order to create a random surface normal structure on a flat surface. With this approach the possible mapping space is sparsely covered. Furthermore we created six different material datasets with different gloss values using a mixture of the Diffuse BSDF and Glossy BSDF node shaders. In this model two parameters can be changed, namely the gloss factor (controlled by the mix node) and the roughness of the two BSDF nodes. For the sake of simplicity, the roughness is the same for both, the Diffuse and the Glossy BSDF node. A schematic illustration of this setup can be seen in Fig. 4. This model is based on a presentation from Gastaldo [17], where he states: R+T+A=1 (3) whereR denotes reflectivity,T denotes transparency and A denotes absorption. Furthermore he states that reflectivity can be divided into diffuse reflectivity (Rd) and specular reflectivity (Rs). With this he derives: Rd+Rs+T ≤1 (4) In our setupRd correlates to the Diffuse BSDF node and Rs to the Glossy BSDF node. Transparency was not taken into consideration (i.e. is always 0) as we exclude glass like materials from our data. The roughness parameter of the Diffuse BSDF node corresponds with the roughness of the Oren-Nayar reflectance model [18]. The model used for the glossy factor of the material was GGX [19]. The roughness parameter of the GGX model simulates microscopic bumps in the surface, so that the reflections of the material look blurrier the higher the roughness parameter is. We excluded a glossy dataset with a roughness value of 0, which would imitate a mirror like behavior. However, 154