Abstract
The University of Southampton (UK) and York University (Canada) have collaborated to build the Southampton-York Natural Scenes (SYNS) public dataset. Our goal is to provide a resource that can be used to relate properties of the human visual system to the statistics of natural scenes. At each scene a 3D point cloud was captured with a Leica P20 LiDAR system over a nearly spherical field of view. Registered spherical high dynamic range monocular imagery and panoramic stereo pairs were also recorded. To derive statistical models of natural surfaces, and to relate these surface models to photometric information in associated imagery, we must first develop reliable methods for estimating surface properties from point cloud data. A standard method for identifying the surface normal at a selected 3D point is to compute the smallest eigenvector of the spatial covariance matrix of k points lying closest to selected point. The main problem with this approach is determining the optimal neighbourhood size k. Here we evaluate a novel adaptive method based upon leave-one-out cross-validation. Specifically, at each point we sweep over a broad range of potential neighborhood sizes (k = 4…100), each time computing k estimates of the surface normal based on k-1 points and measuring error as the deviation of the remaining validation point from the estimated tangent plane. The optimal k is that which yields the lowest mean k-fold cross-validation error. We demonstrate that this adaptive method works reliably for diverse scenes in urban and rural environments. Typically a very local neighbourhood (mode of k = 8) is selected, but the distribution has a strong positive tail, particularly for urban environments, where planar surfaces can be estimated more reliably with larger neighbourhoods. We discuss how higher-order geometric surface properties of natural scene data can estimated using similar methods.
Meeting abstract presented at VSS 2015