URSA: A Neural Network for Unordered Point Clouds Using Constellations
share
This paper describes a neural network layer, named URSA, that uses a constellation of points to learn classification information from point cloud data. Similar to new methods such as PointNet, this architecture works directly on D-dimensional points rather than first converting the points to a D-dimensional volume. URSA is invariant to permutations of the input data. We use an URSA layer, followed by a series of dense layers, to classify 2D and 3D objects from point cloud. Experiments on ModelNet40 and MNIST data show classification results comparable with current methods, while requiring half or fewer trained parameters.
READ FULL TEXT