Discussion on Image Classification of Vehicle Make and Model Using Convolutional Neural Networks and Transfer Learning

         In this article, the authors have tried to explain image classification done for vehicle make & models with the help of convolutional neural networks, as well as, their transfer learning. It has been said that the identification and detection of cars is one of the major tasks with regard to traffic management and control. To handle these kinds of tasks, a variety of domain-specific features along with large datasets are used so that data can better fit-in to get considerable results. In this study, the authors have tried to train, implement, as well as, test some of the best classifiers, which are trained to be used with domain-general datasets so that make & models of the cars can be identified properly.

            So, they decided to do an experiment with transfer learning at different levels so that it can be fit for the given models. The cars come with different make and models so computers are not able to identify cars like humans. So, better methods are needed for the identification of cars. The authors used a variety of data sets and sources to continue with the experiment so that viable findings are made. After completing the experiment with given methods and sources, it was found that transfer learning was extremely crucial to get great performance, as the amount of data was small, whereas the number of classes was large. This experiment outperformed the nets, which were trained from randomly initialized, as well as, small weights.

Article Link: http://cs231n.stanford.edu/reports/2015/pdfs/lediurfinal.pdf