Analyzing Most Popular Object Detection Models for Deep Neural Networks

With the rapid evolution of deep convolutional neural networks (CNNs), major breakthroughs have been achieved in object detection in the field of computer vision. However, the majority of state-of-the-art detectors, in both one-stage and two-stage methods, have limits and are inadequate for usage in a real-world setting where each step must be thoroughly checked. This thesis investigates advanced object detection models and frameworks. It offers an in-depth analysis of the most recent object detection models, their frameworks, and the performance criteria used to evaluate such models. The object detection models selected are YOLOv5, Faster R-CNN using Detectron 2, and SSD using TensorFlow 2, and the dataset selected is the Vehicles-Open Images Dataset. The performance of the selected models in relation to many metrics is analyzed, and the findings are reported. In conclusion, the benefits and limits of the selected models, as well as their relative performance, are discussed.