Fine Tuning Based SqueezeNet for Vehicle Classification

The vehicle classification contributes in many important aspects of road management. In this paper, we convey our result of the vehicle classification based on the deep net prediction. We fine-tuned the recently published SqueezeNet on the Miovision Traffic Camera Dataset (MIO-TCD)1. The imbalanced instance distribution of the MIO-TCD classification training section portrays the nature of traffic scene. To give a better generalization, we redesign the instance distribution of the training dataset by implementing augmentation and reduction. We find that the balanced instance distribution delivers significant minority class prediction accuracy improvement over the imbalanced instance training dataset. Also, we show that fine-tuning a pre-trained SqueezeNet effectively increase the net performance over the baseline models.