BUILDING 3D SHAPE PRIMITIVE BASED OBJECT MODELS FROM RANGE IMAGES
Abstract
Most pattern recognition approaches to object identification work in the image
domain. However this is ignoring potential information that can be provided by depth
information. Using range images, we can build a set of geometric depth features. These
depth features can be used to identify basic three-dimensional shape primitives.
There have been many studies regarding object identification in humans that
postulate that at least at a primary level object recognition works by breaking down
objects into its component parts. To build a similar Recognition-by-component (RBC)
system we need a system to identify these shape primitives.
We build a depth feature learner by extending a sparse autoencoder neural
network into a model similar to a convolutional neural network to learn supersized
features that can be matched to patches extracted from depth images. This allows us to
convert a collection of patches from a depth image of an object into converted into the
space defined by the best fit on each of these supersized features. We also train a
backpropagation network to identify shape primitives from patches from known shape
primitives that have been converted into this feature space.