Page - 71 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Depth estimation using light fieldsandphotometric stereo with amulti-line-scanframework DorisAntensteiner1, Svorad Sˇtolc1, andReinholdHuber-Mo¨rk1 Austrian InstituteofTechnology DigitalSafetyandSecurityDepartment, Austria {doris.antensteiner.fl,svorad.stolc,reinhold.huber-moerk}@ait.ac.at Abstract In thispaperwedealwithacombinationof twostate-of-the-art computational imagingapproaches - (i) lightfieldsand(ii)photometricstereo- inordertoimprovethequalityof3Dreconstructionswithin amulti-line-scanframework. Computational imagingusesaredundantdescriptionofanimagescene to reveal information which would not have been available via conventional imaging techniques. In thecaseof lightfields the redundancy isachievedbyobserving thescene frommanydifferentangles, which allows capturing 3D shapes in areas with a prominent surface structure using stereo vision techniques. Contrarily, photometry makes use of multiple illuminations in order to capture local sur- face deviations without the necessity of any surface structure. As photometric surface reconstruction is very sensitive to fine surface details and light fields excel in capturing global shapes, naturally a more complete description can be achieved through a combination of both techniques. We present a compact hybrid photometric light field setup with relatively low costs and improved accuracy, which is therefore well suited for industrial inspection. A multi-line-scan camera is statically coupled with an illumination source to obtain light field data which is also comprised of photometric information. Novel algorithms have been developed to use this data for an improved 3D reconstruction, which exhibits large-scaleaccuracyaswell as sensitivity to finesurfacedetails. 1. Introduction Traditional film cameras as well as digital cameras capture light rays and project images of the envi- ronmentontoa2Dplane. Non-traditionalapproachessuchasmulti-cameraarrays,plenopticcameras or coded apertures capture a portion of the so called 4D light field and further subdivide this ray space with respect to position and orientation. Photometric stereo makes explicit use of directional variation of illumination. The presence of different lighting conditions, in cases where the light field is dynamically constructed over time, introduces a photometric variation into the data structure. The combination of light fields, describing the variation of image content over observational directions, and photometric approaches, describing the variation of image content depending on lighting direc- tions, is apromising researchdirection. In general, the combination of methods which are (i) locally precise but globally inaccurate with (ii) globallyaccuratemethods,whichare lacking local structure,wasapproachedfromdifferentperspec- tives, e.g. combining shading with RGB-D [12], improving a depth map from time-of-flight with polarization cues [3], or using stereo vision with photometric stereo [5]. Methods used for the com- binationareeitherbasedon the fusionofnormalvectorsprovidedbydifferentapproaches, the fusion ofdepthmeasurements, or employingvariationalmethods. 71