Unsupervised Feature Learning for Classification of Outdoor 3D Scans
Feature learning on dense 3D data has proven to be a highly successful alternative to manual hand crafting of features. The data produced by outdoor 3D scanning devices is typically unsuitable for feature learning approaches due to its variable density. We introduce a method for formatting segmented regions of Velodyne scans into regularly sampled depth images. This reformatting allows us to leverage existing feature learning techniques and apply them to Velodyne data. Experiments are performed on a set of 588 object scans labelled into 14 categories. Features are learnt on the raw Velodyne range images and evaluated on the proposed interpolated depth images. Classification performance produced by learnt features compare favourably against classifiers based on pre-defined 3D features such as the Spin Images or other more recent alternatives.
PDF
Sydney Urban Objects
