Open Source Assessment of Deep Learning Visual Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models for visual object detection require robust evaluation methods.
• Comparing models across different frameworks and datasets presents significant challenges.
• Objective and standardized performance metrics are crucial for advancing the field.

Purpose of the Study:

• To introduce Detection Metrics, an open-source software tool for assessing deep learning neural network models in visual object detection.
• To provide objective performance metrics and facilitate fair comparisons between diverse models and datasets.
• To offer tools for dataset management, model visualization, and batch processing to streamline research workflows.

Main Methods:

• The software supports major international object detection datasets and popular deep learning frameworks.
• It calculates objective performance metrics including mean average precision and mean inference time.
• Includes utilities for dataset/model management, format conversion, and visualization.

Main Results:

• Detection Metrics enables fair comparison of different network models, irrespective of their framework.
• The software has been experimentally validated and used in research projects to guide model selection.
• Facilitates automatic batch processing for large-scale experiments and creation of custom datasets.

Conclusions:

• Detection Metrics provides a valuable, open-source solution for the objective evaluation of visual object detection models.
• Its comprehensive features save researchers time and enhance the reproducibility of deep learning experiments.
• The tool supports the development and selection of optimal deep learning architectures and training strategies.