3DMatch: Learning Local Geometric Descriptors from RGB-D Reconstructions APPENDIX

As mentioned in Sec. 3 of the main paper, we use registered depth frames of 62 different real-world scenes collected from Analysis-by-Synthesis [6], 7-Scenes [5], SUN3D [7], RGB-D Scenes v.2 [4], and Halber et al. [3], with 54 scenes for training and 8 scenes for testing. For selected scenes of the SUN3D dataset, we use the method from Halber et al. to estimate camera poses. For the precise train/test scene splits, see our project webpage [1]. In Fig. 2, we show top-down views of the completed reconstructions. They are diverse in the environments they capture, with local geometries at varying scales, which provide a diverse surface correspondence training set for 3DMatch. In total, there are 214,569 depth frames over the 54 training scenes, most of which are made up of indoor bedrooms, offices, living rooms, tabletops, and restrooms. The depth sensors used for scanning include the Microsoft Kinect, Structure Sensor, Asus Xtion Pro Live, and Intel RealSense. The size of the correspondence training set correlates with the amount of overlap between visible surfaces from different scanning views. In Fig. 1, we show the average distribution of volumetric voxels (size 0.023m) on the surface vs. the number of frames in which the voxels were seen by the depth sensor. We plot this distribution averaged over the 54 training scenes (left) and illustrate a heat map of over two example reconstructions (right), where a warmer region implies that the area has been seen more times. The camera trajectories are plotted with a red line.