Recognizing Actions of Distracted Drivers using Inception v3 and Xception Convolutional Neural Networks

Abstract

In recent years, Deep Convolutional Neural Networks have shown remarkable success in image classification tasks. In our research, we compare the performance of two competing DCNN architectures, viz. Inception v3 and Xception, and use them to recognize ten unique actions of the drivers in the Kaggle’s State Farm Distracted Driver Detection challenge. We discuss the performance of both architectures in detail (in terms of loss, accuracy and Kaggle test set scores) under two different weight initialization schemes, i.e. random initialization and transfer learning using ImageNet weights, with the training set split on drivers. Additionally, for comparison, we also used a randomly split training set and trained the models using ImageNet weights. By splitting training dataset on drivers, we find that high top-1 validation accuracy of 85.4% is achieved by training the Xception architecture using transfer learning with ImageNet initialized weights. This accuracy is further increased to 99.3% for the same Xception architecture scheme, when we split training data randomly instead of splitting it on subjects. Our best trained model utilizing the Xception architecture with ImageNet initialized weights ranks at 325th position (out of 1440 entries) on Kaggle’s Private Leaderboard with a remarkable test loss of 0.51285.