Labelling images with a neural network

Shows that a neural network with a multilayer perceptron architecture is capable of automatically labelling the visible objects in colour images of outdoor road scenes. The two problems of segmentation and recognition are separated by using `ideal´ segmentations, allowing the performance of the recognition method to be studied independently of the effects of using an imperfect real segmentation process. The authors argue that they have sufficient training data to constrain the degrees of freedom in the network. A label clustering transformation is proposed which results in a significant increase in the expected classification accuracy of the network. They demonstrate the importance of the contextual features with a control experiment in which the loss of the contextual features is shown to degrade the performance of the re-trained network