Open-Source Object Detection Frameworks : A comparative analysis

Sammanfattning

This thesis examines the usability and accessibility of open-source object detection frameworks, focusing on their suitability for programmers new to computer vision. Popular frameworks such as Ultralytics, MMDetection, and PaddleDetection – among others – were evaluated based on ease of installation, documentation quality, abstraction levels, features, inference speed benchmarks and time required to install and implement a standardised project. Results indicate that Ultralytics stands out for its user-friendly interface, documentation, and thorough support for YOLO models. MMDetection offers a broader selection of features and models but poses significant setup challenges. PaddleDetection provides extensive model coverage but has limited English documentation, Detectron2 is generally high-quality but limited to R-CNN models, while Supervision only offers utility functions without inference support, making them less suitable for general recommendation. Installation issues and bugs prevented full evaluations of the Detrex and PAZ frameworks; however, these were similarly limited in their model selection. The findings provide actionable insights for first-time developers, recommending tools that balance usability and functionality to lower entry barriers in computer vision projects, improve reproducibility and encourage academic adoption. Additionally, they highlight areas where frameworks can improve their design to better support novice users.