A Deep Learning Framework Performance Evaluation to Use YOLO in Nvidia Jetson Platform

Round 1

Reviewer 1 Report

This paper presents an empirical evaluation of using YOLO in embedded environments. The authors use one data collection for assessing four deep learning frameworks. It is hard to predict the behavior of packages when they are explored on just one collection. I recommend describing more the data characteristics, and to provide a statistical investigation of the results on more data collections.

Author Response

We appreciate your comments. We believe that the comments have helped us improve the quality of our paper. Once again, thank you for your time and effort and the opportunity to improve our paper.

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this paper authors perform comparative study of four software platforms that implement deep learning algorithms, namely TensorFlow, TensorFlow-Lite, TensorFlow-TensorRT, and TensorRT. The first two developed by Google, and the latter two optimized for Nvidia hardware. As an exemplar problem the authors selected object detection problem with implementation of YOLOv4 detector. The detector was implemented on hardware system, produced by Nvidia, Jetson AGX Xavier 16GB.

The paper provides structured review of prior works in the area of object detection and YOLO development, hardware comparison of different models of Nvidia Jetsons, and platforms comparison, that implement deep learning algorithms for different devices. The algorithm implementations on different software platforms are compared in terms of Average CPU utilization, Average GPU utilization, Average Latency, and Average Power. At the end authors provide a comprehensive conclusions what software platform to use for Nvidia Jetson AGX Xavier 16GB.

The paper is well written however, contains minor flaws: in line 239 it should read "...can expOrt regardless...", instead of "...can expErt regardless...". Also in line 242: "... regardless of ... or flat" it is not clear what word "flat" stands for.

Author Response

We appreciate your comments. We believe that the comments have helped us improve the quality of our paper. Once again, thank you for your time and effort and the opportunity to improve our paper.

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

In the abstract make clear what your contributions are it is really difficult to pick out what is novel about this research... 

In 3.3 you state some contributions but not the novelty of this research other than the comparative analysis of the hardware. -is this sufficient?

The scope of this paper is really limited,

 

'Through this study, it is expected to present guidelines on which deep learning framework should be used when researching using Jetson products or developing applications based on them.'

So this paper is only going to be of interest to people using Jetson - Nvidia products? This is a very limited audience and should be reflected in the title of the paper, as by the time the reader hits this section in the middle of the paper they realise, oh I need to be using this product for this paper to mean anything.

This paper is very convoluted, although it reads well it is really difficult to follow. Even in the conclusion it's difficult to understand what has actually been done here other than a comparative analysis - and even then, is this novel or contribute to anything outside of Jetson - Nvdia products?

On this basis, I think this paper needs some serious consideration, make it clear to the reader what your contributions are, the novelty of this research and what audience it is aimed at.

 

 

Author Response

We appreciate your comments. We believe that the comments have helped us improve the quality of our paper. Once again, thank you for your time and effort and the opportunity to improve our paper.

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

In my opinion, the authors treated correctly all the issues from the first revision.

Reviewer 3 Report

With the amendments made the paper is now acceptable.

Although the scope of the paper is limited due to the hardware, as this is now reflected in the title.. I believe it would be of interest to that specific community.