Unified Framework for Vision Inference on the Edge
Abstract
Edge processing for computer vision systems enable incorporating visual
intelligence to mobile robotics platforms. Demand for low power, low
cost and small form factor devices are on the rise.This work proposes a
unified platform to generate deep learning models compatible on edge
devices from Intel, NVIDIA and XaLogic. The platform enables users to
create custom data annotations,train neural networks and generate edge
compatible inference models. As a testimony to the tools ease of use and
flexibility, we explore two use cases --- vision powered prosthetic hand
and drone vision. Neural network models for these use cases will be
built using the proposed pipeline and will be open-sourced. Online and
offline versions of the tool and corresponding inference modules for
edge devices will also be made public for users to create custom
computer vision use cases.