Our 3D Vision Data-Sets in the Making

Over the previous years, we have at the Section for Image Analysis and Computer Graphics at the Technical University of Denmark been working on generating high quality data sets for computer vision via our lab setup using a 6-axis industrial robot. This has provided a new data set aimed at feature matching [1, 4], and two data sets aimed at multiple view stereo [16, 18]. The resulting data sets are publicly available via http://roboimagedata. compute.dtu.dk/. The evaluation of computer vision algorithms on these data sets has provided useful insights on realistic scenarios by setting a rigorous framework for evaluation. The results of these efforts have been well received by the community and the hardware and software platform associated with the robot is now well developed. We are currently in the process of making three new data sets aimed at 3D vision, with a special focus on the more challenging aspects, such as radiometry and the modelling of non-rigid objects. The construction of these data sets all leverage on our robotic setup’s ability to produce ground truth camera and surface geometry, as briefly outlined in Section 2, and there is a great deal of commonality in the making of the data sets. This abstract describes our current ongoing work on this data set construction for 3D vision. The data sets include: