DAIDESS - Decomposable AI Deployments made Efficient and Sustainable by Specialization

Purpose and goal:

Artificial Intelligence (AI) and Machine Learning (ML) have rapidly moved from academic research to practical applications. For deployments, there are various accelerators, both in the cloud and in edge devices, but availability varies. Our project aims to improve the usability of computer vision algorithms and platforms by focusing on three aspects: research on generic, decomposable computer vision algorithms, platform development to deploy these algorithms, and address a specific use case in automated sports production.

Expected results and effects:

The project will enable, simplify, and automate the decomposition and adaptation of computer vision algorithms that can lead to deep learning solutions implementable in a more economically profitable and environmentally friendly way for different forms of hardware. We aim for a versatile platform for various industrial purposes, but we will show this by implementing a case of technical analysis of video from team sports in cameras, cloud and display devices.

Approach and implementation: The project mainly follows three tracks with synergies:

(1) Building a platform that can dynamically partition and schedule algorithms on currently available efficient deep learning processors or accelerators. (2) To develop general enabling technologies and methods to easily produce decomposable algorithms. (3) Specialize general algorithms to the specific application/scene/ of interest using known prerequisites.