Optimization of Computer Vision Software Models for Deployment to Performance Constrained Embedded Processors
- 提供方法
- 版元よりダウンロードリンクを連絡
- 形態
- 価格
- 一般価格(税込)：￥6,600 会員価格(税込)：￥5,280
- 文献・情報種別
- SAE Paper
No.2022-01-0160
- 掲載ページ
- 1-9(Total 9 p)
- 発行年月
- 2022年 3月
- 出版社
- SAE International
- 言語
- 英語
- イベント
- WCX SAE World Congress Experience 2022
書誌事項
| 著者（英） | 1) Davis Sawyer, 2) Benjamin M. Rocci, 3) Sudhakar Sah, 4) Ehsan Saboori, 5) Olivier Mastropietro |
|---|---|
| 勤務先（英） | 1) Deeplite, 2) Deeplite, 3) Deeplite, 4) Deeplite, 5) Deeplite |
| 抄録（英） | Computer vision (CV), a form of artificial intelligence (AI), is a foundational technology within the automotive industry for an increasing number of applications including active safety, motion control, and driver distraction monitoring. State-of-the-art CV models often rely on the use of Deep Neural Networks (DNNs) to achieve high levels of accuracy. While necessary for their accuracy, DNNs are computationally complex. For example, when compared to other AI model architectures, they have a large memory footprint and often require billions of floating-point operations to create an output or prediction. To meet performance goals in the face of such constraints, high performance processors such as Graphics Processing Units (GPUs) are typically required to run CV models on-board automobiles, creating a major hurdle to the deployment of CV applications. This paper proposes and analyzes a method for optimizing and compressing DNN-based CV models to enable their deployment to lower performance processors - such as Central Processing Units (CPUs) - more commonly used in automotive applications. Several real-world case-studies are put forward which demonstrate the effectiveness of the method including a seventeen-fold reduction in model size and a three-fold reduction in inference latency for AI-powered object detection on an automotive CPU. 翻訳 |
