In the last decades, image processing has moved from academic research to
innovative consumer applications. One of the most valuable of these practical
uses is in object detection: from an image, identify and locate elements.
One family of object detectors based on deep learning is known as YOLO.
The goal of this work is to benchmark an object detection model based on
YOLO version 5 regarding detection metrics as well as timing and energy
consumption. We evaluate the performance of a vehicle license plate detector
as well as examine what software frameworks and hardware resources are
most suitable for the task. We present the literature about object detection
and discuss performance metrics. We describe the dataset of license plates
curated for the experiments and the training procedure. Then, we present
performance results for different software stacks, with PyTorch as the baseline,
and hardware equipment, CPUs and GPUs part of the DaVinci-1 cluster
at Leonardo SpA and Intel Developer Cloud Beta. Besides that, energy consumption
results are also discussed. Finally, we evaluated the effect of larger
image sizes on the inference time as well as of grouping images in the same
batch for processing.
