Page - 73 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

viewpoint alongonedirection. Figure2: DragonStanford[13]scene image(a), zoomedinat the reddotted line (b),whereEPIstacks with9differentviewpoints are shown, each stack formedwithacertain lightdirection (c). We analyse the captured light fields in the EPI domain [1]. A cut through the light field stack shows linearslopestructures,where theangleof theslopecorresponds to thedisparityandtherebythedepth of the scene, as shown in Fig. 2. Each angle of a slope in the EPI stack corresponds to a defined distance between the camera and the object point. Photometric information is obtained by a static light source w.r.t. the sensor while the object is moving. As an object moves on the conveyor belt, the relationship between the illumination and the observation angle changes in a systematic way, so that the surface inclination in the transport direction can be estimated. This photometric information isused toestimate the surfacenormalsof theobject. 3. Combinationof LightFieldsandPhotometricStereo Photometric stereodescribes thesurfacevariationw.r.t. the lightingdirection. Reflectionsof the light on the objects’ surface under different lighting orientations provide information about the surface normals at each object point. We combine the depth from light fields with fine surface structures as observed by photometric stereo, to gain an improved depth map of the scene. Figure 3 shows the depthestimationachievedusingboth lightfieldandphotometric stereo independently inavirtual test setup,wherewesimulated81cameraviewpoints and25 illuminationangles. 3.1. LightField Depth Estimation Depth information of a scene can be retrieved by analyzing the slope angles in EPI stacks. An EPI slice of this stack is shown in Fig. 2, where each angle in the slice refers to a defined depth of a corresponding point in the scene. Using this data we gain a rough absolute depth estimation of the scene. Analyzing the depth from EPI stacks can be seen as finding such an angleα∗ for each point (x,y), where the difference between values at the sheared coordinates (x(α),y(β)) in the light fieldL and the referenceviewI0 isminimal [7]. α∗(x,y)=argmin α,β i=n∑ i=1 |Ii(x(α),y(β))−I0(x,y)| The use of a block-matching approach creates a higher robustness to both noise and non-Lambertian lightingconditions. AsshowninFig.3b, lightfielddataprovidesquite robust absolutedepthestima- tion,but lacksprecision infinesurfacedetails. 73