COMPARATIVE ANALYSIS OF OBJECT DETECTION PROCESSING SPEED ON THE BASIS OF NEUROPROCESSORS AND NEUROACCELERATORS

Аннотация

This paper is devoted to a comparative analysis of neural network models based on neuroprocessors.
The following neural network models were selected: MobileNetSSD v1, SSD MobileNet v2. As a research task, the
authors determined an attempt to compare several platforms that differ in size, computational capabilities and cost:
Coral Dev Board, NVIDIA Jetson Nano, Coral USB Accelerator, Neural Compute Stick 2, Raspberry Pi 4. Local
data processing offers a number of advantages compared to downloading calculations to a remote server or data
center. Firstly, downloading data to remote servers takes a lot of time, as well as additional costs for infrastructure
with energy, financial and computer equipment. It also requires high bandwidth and reliability, as data transfer may
not be completed in case of a bad signal. Secondly, data transfer can lead to security and privacy issues. Finally, local
processing can reduce the amount of data transferred to the cloud, which allows to performing tasks of a higher level.
The aim of this paper is a comparative analysis of platforms for object recognition tasks (object detection) with
MobileNetSSD v1 / v2 models. For training, a cloud service based on the Jupyter Notebook, which gives access to
incredibly fast GPUs and TPUs was used. The paper addresses the topic of determining small objects using the
example of car detection.
Based on the study of platforms and models, it was found that the MobileNetSSD v1 model is effective for
NVIDIA Jetson Nano by NVIDIA (61FPS), but in turn, the MobileNet v2 SSD model is less efficient (11FPS).
Google’s Coral Dev Board is more productive than other devices (47.8FPS and 63FPS). Raspberry Pi 4 (0.8FPS and
1.4FPS) turned out to be less effective. Among neuroprocessors, Neural Compute Stick 2 (9FPS and 7.1FPS) showed
poor performance.