Comparative analysis of neural network models performance on low-power devices for a real-time object detection task
A computer vision based real-time object detection on low-power devices is economically attractive, yet a technically challenging task. The paper presents results of benchmarks on popular deep neural network models, which are often used for this task. The results of experiments provide insights into...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Samara National Research University
2024-04-01
|
| Series: | Компьютерная оптика |
| Subjects: | |
| Online Access: | https://www.computeroptics.ru/eng/KO/Annot/KO48-2/480211e.html |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1832540917870886912 |
|---|---|
| author | A. Zagitov E. Chebotareva A. Toschev E. Magid |
| author_facet | A. Zagitov E. Chebotareva A. Toschev E. Magid |
| author_sort | A. Zagitov |
| collection | DOAJ |
| description | A computer vision based real-time object detection on low-power devices is economically attractive, yet a technically challenging task. The paper presents results of benchmarks on popular deep neural network models, which are often used for this task. The results of experiments provide insights into trade-offs between accuracy, speed, and computational efficiency of MobileNetV2 SSD, CenterNet MobileNetV2 FPN, EfficientDet, YoloV5, YoloV7, YoloV7 Tiny and YoloV8 neural network models on Raspberry Pi 4B, Raspberry Pi 3B and NVIDIA Jetson Nano with TensorFlow Lite. We fine-tuned the models on our custom dataset prior to benchmarking and used post-training quantization (PTQ) and quantization-aware training (QAT) to optimize the models’ size and speed. The experiments demonstrated that an appropriate algorithm selection depends on task requirements. We recommend EfficientDet Lite 512×512 quantized or YoloV7 Tiny for tasks that require around 2 FPS, EfficientDet Lite 320×320 quantized or SSD Mobilenet V2 320×320 for tasks with over 10 FPS, and EfficientDet Lite 320×320 or YoloV5 320×320 with QAT for tasks with intermediate FPS requirements. |
| format | Article |
| id | doaj-art-c6f4cbee7d484381910b3ec57d741bb0 |
| institution | Kabale University |
| issn | 0134-2452 2412-6179 |
| language | English |
| publishDate | 2024-04-01 |
| publisher | Samara National Research University |
| record_format | Article |
| series | Компьютерная оптика |
| spelling | doaj-art-c6f4cbee7d484381910b3ec57d741bb02025-02-04T12:44:09ZengSamara National Research UniversityКомпьютерная оптика0134-24522412-61792024-04-0148224225210.18287/2412-6179-CO-1343Comparative analysis of neural network models performance on low-power devices for a real-time object detection taskA. Zagitov0E. Chebotareva1A. Toschev2E. Magid3Institute of Information Technology and Intelligent Systems, Kazan Federal UniversityInstitute of Information Technology and Intelligent Systems, Kazan Federal UniversityInstitute of Information Technology and Intelligent Systems, Kazan Federal UniversityInstitute of Information Technology and Intelligent Systems, Kazan Federal University; School of Electronic Engineering, Tikhonov Moscow Institute of Electronics and Mathematics, HSE UniversityA computer vision based real-time object detection on low-power devices is economically attractive, yet a technically challenging task. The paper presents results of benchmarks on popular deep neural network models, which are often used for this task. The results of experiments provide insights into trade-offs between accuracy, speed, and computational efficiency of MobileNetV2 SSD, CenterNet MobileNetV2 FPN, EfficientDet, YoloV5, YoloV7, YoloV7 Tiny and YoloV8 neural network models on Raspberry Pi 4B, Raspberry Pi 3B and NVIDIA Jetson Nano with TensorFlow Lite. We fine-tuned the models on our custom dataset prior to benchmarking and used post-training quantization (PTQ) and quantization-aware training (QAT) to optimize the models’ size and speed. The experiments demonstrated that an appropriate algorithm selection depends on task requirements. We recommend EfficientDet Lite 512×512 quantized or YoloV7 Tiny for tasks that require around 2 FPS, EfficientDet Lite 320×320 quantized or SSD Mobilenet V2 320×320 for tasks with over 10 FPS, and EfficientDet Lite 320×320 or YoloV5 320×320 with QAT for tasks with intermediate FPS requirements.https://www.computeroptics.ru/eng/KO/Annot/KO48-2/480211e.htmlcomputer visionimage analysisobject detectiondeep learningbenchmarkingoptimization techniques |
| spellingShingle | A. Zagitov E. Chebotareva A. Toschev E. Magid Comparative analysis of neural network models performance on low-power devices for a real-time object detection task Компьютерная оптика computer vision image analysis object detection deep learning benchmarking optimization techniques |
| title | Comparative analysis of neural network models performance on low-power devices for a real-time object detection task |
| title_full | Comparative analysis of neural network models performance on low-power devices for a real-time object detection task |
| title_fullStr | Comparative analysis of neural network models performance on low-power devices for a real-time object detection task |
| title_full_unstemmed | Comparative analysis of neural network models performance on low-power devices for a real-time object detection task |
| title_short | Comparative analysis of neural network models performance on low-power devices for a real-time object detection task |
| title_sort | comparative analysis of neural network models performance on low power devices for a real time object detection task |
| topic | computer vision image analysis object detection deep learning benchmarking optimization techniques |
| url | https://www.computeroptics.ru/eng/KO/Annot/KO48-2/480211e.html |
| work_keys_str_mv | AT azagitov comparativeanalysisofneuralnetworkmodelsperformanceonlowpowerdevicesforarealtimeobjectdetectiontask AT echebotareva comparativeanalysisofneuralnetworkmodelsperformanceonlowpowerdevicesforarealtimeobjectdetectiontask AT atoschev comparativeanalysisofneuralnetworkmodelsperformanceonlowpowerdevicesforarealtimeobjectdetectiontask AT emagid comparativeanalysisofneuralnetworkmodelsperformanceonlowpowerdevicesforarealtimeobjectdetectiontask |