Andrew Rabinovich, Director of Deep Learning at Magic Leap, presents the "Performing Multiple Perceptual Tasks With a Single Deep Neural Network" tutorial at the May 2017 Embedded Vision Summit.
As more system developers consider incorporating visual perception into smart devices such as self-driving cars, drones and wearable computers, attention is shifting toward practical formulation and implementation of these algorithms. Here, the key challenge is how to deploy very computationally demanding algorithms that achieve state-of-the-art results on platforms with limited computational capability and small power budgets. Visual perception tasks such as face recognition, place recognition and tracking are traditionally solved using multiple single-purpose algorithms. With the approach, power consumption increases as more tasks are performed.
In this talk, Rabinovich introduces techniques for performing multiple visual perception tasks within a single learning-based algorithm. He also explores general-purpose model optimization to enable such algorithms to run efficiently on embedded platforms.