Abstract—Vehicle identification via computer vision is an important task for traffic control, video surveillance, and security and authentication. However, there are challenges with image classification of vehicles due to the fine-grained details that are inherently harder for a computer to detect. Using limited data from the Stanford Cars dataset, we implement transfer learning on a pre-trained Convolutional Neural Network (CNN) framework to classify vehicles based on 196 classes of different vehicle makes, models, and years. After fine-tuning, we achieve a test accuracy of 85% and top-5 test accuracy of 96.3%, surpassing state-of-the-art results.
Keywords—Vehicle identification, computer vision, fine-grained details, transfer learning, Convolutional Neural Network
Full paper here.
Github code here.