Three-Label Outdoor Scene Understanding Based on Convolutional Neural Networks

Scene understanding is the task of giving each pixels in an image a label, which is the class of the pixel belongs to. Traditional scene understanding is object-based approach, which has lots of limitations as the descriptors cannot give the whole characteristics. In this paper, a convolutional neural network based method is proposed to extract the internal features of the whole image, then a softmax regression classifier is applied to generate the label. Scene understanding used in self-navigating vehicles only concentrate on the road, so the number of classes is reduced in order to get higher accuracy by lower computational cost. A preprocessing is implemented on Stanford Background Dataset to obtain three-label images including road, building, and others. As a result, the system yields high accuracy on the three-label dataset with great speed.