View Sphere Partitioning via Flux Graphs Boosts Recognition from Sparse Views

View-based 3D object recognition requires a selection of model object views against which to match a query view. Ideally, for this to be computationally efficient, such a selection should be sparse. To address this problem, we partition the view sphere into regions within which the silhouette of a model object is qualitatively unchanged. This is accomplished using a flux-based skeletal representation and skeletal matching to compute the pairwise similarity between two views. Associating each view with a node of a view sphere graph, with the similarity between a pair of views as an edge weight, a clustering algorithm is used to partition the view sphere. Our experiments on exemplar level recognition using 19 models from the Toronto Database and category-level recognition using 150 models from the McGill Shape Benchmark demonstrate that in a scenario of recognition from sparse views, sampling model views from such partitions consistently boosts recognition performance when compared against queries sampled randomly or uniformly from the view sphere. We demonstrate the improvement in recognition accuracy for a variety of popular 2D shape similarity approaches: shock graph matching, flux graph matching, shape context-based matching, and inner distance-based matching.