A benchmark dataset and methodology for fine grained vehicle make and model classification

Urban transportation management increasingly relies on Intelligent Transportation Systems (ITS), where Vehicle Make and Model Recognition (VMMR) plays a vital role in surveillance, traffic monitoring, and infrastructure planning. However, traffic conditions in developing nations such as Pakistan present unique challenges due to unstructured driving practices and lack of lane discipline. We introduce a large VMMR dataset for Pakistan's traffic dynamics to address these challenges. This dataset comprises 129,000 images across 94 vehicle classes. We collected the dataset through web scraping and overhead traffic video recording, followed by an iterative semi-automated annotation process to ensure quality and reliability. For evaluation, we perform a fine-grained analysis using modern deep-learning architectures, including VGG, EfficientNet, and Vision Transformers. Experimental results are obtained through model simulations. These results establish a new benchmark in vision-based traffic analytics for developing countries. Our best-performing model achieves an accuracy of 97.3%, demonstrating the potential of the data set to advance ITS applications.