Abstract
The Southampton-York Natural Scenes (SYNS) dataset consists of LiDAR range and image data captured from a variety of natural and built environments. One of our goals is to use the dataset to relate the ecological statistics of 3D surfaces to human perception of surface attitude. A local planar fit at a 3D point in the dataset can be estimated from an eigen-decomposition of a k-neighbourhood of surrounding points. One challenge is to determine the optimal local scale, k; smaller scales produce noisier estimates, while larger scales lead to over-smoothing. We designed and evaluated two algorithms for adaptively selecting the optimal local scale. The first algorithm uses unsupervised leave-one-out cross-validation (XVAL). For each scale k we fit k planes using k-1 points and measured error of fit as the average distance of the left-out point from the plane. The XVAL method assumes white sensor noise. However, in many sensors, including our own Leica P20, internal post-processing produces correlated noise. To address this problem, we evaluated a second, supervised method based on null hypothesis testing (NHT). Under the NHT approach, the surface is assumed to be locally planar unless there is strong evidence to the contrary. In the NHT training phase, we used a planar reference surface to measure the maximum observed mean deviation of points from fitted planes as a function of scale k and distance. In the estimation phase, this function provides an upper bound on the expected deviation for a locally planar surface; we select the maximum k that does not exceed this bound. Both methods tend to select smaller scales for bumpy surfaces and larger scales for flat surfaces. However, by taking noise correlations into account, the NHT method produces more reliable and more accurate surface attitude estimates for a range of different environments.
Meeting abstract presented at VSS 2016