Fine-To-Coarse Global Registration of RGB-D Scans Supplemental Material

1. Plane All parts of the image that are locally planar, i.e. can be represented as a position and a normal. We construct a planar surface feature for each pixel in coplanar clusters of size ≥ 100. 2. Silhouette We have found it useful to mark depth discontinuities with a separate feature type. These are represented as a point, normal and a direction along the silhouette. Normal and direction are estimated with PCA on neighboring silhouette pixels. The same approach is used to estimate direction for Ridge and Valley features. 3. Ridge If two large planes are intersecting and the angle between their normals is larger than π we mark the intersection as a ridge feature. 4. Valley If two large planes are intersecting and the angle between their normals is smaller than π we mark them as a valley feature. Both valley and ridge features are represented as a position, normal (average of the normals of neighboring planes) and direction along the intersection.