Deep learning techniques are gaining in popularity in many facets of
embedded vision, and this holds true for AR and VR. Will they soon dominate
every facet of vision processing? This talk explores this question by
examining the theory and practice of applying deep learning to real world
problems for Augmented Reality, with real examples describing how this shift
is happening today quickly in some areas, and slower in others.
Today it’s widely accepted for image recognition tasks that Deep Learning
techniques involving Convolutional Neural Networks (CNNs) are dominating.
Other application solutions use hybrid approaches. Other applications are
still holding out using classical embedded vision techniques.
These themes are then explored further through specific real world examples
of gesture tracking (which is moving to CNN), stereo depth (hybrid
approach), SLAM (moving toward a hybrid approach), and ISP (imaging
pipelines holding with traditional algorithms). The talk contrasts the mix
of algorithms being deployed today with a prediction of the mix we expect to
find in AR/VR headsets 3 years from now.