Using deep autoencoders to investigate image matching in visual navigation

This paper discusses the use of deep autoencoder networks to find a compressed representation of an image, which can be used for visual naviga-tion. Images reconstructed from the compressed representation are tested to see if they retain enough information to be used as a visual compass (in which an image is matched with another to recall a bearing/movement direction) as this ability is at the heart of a visual route navigation algorithm. We show that both reconstructed images and compressed representations from different layers of the autoencoder can be used in this way, suggesting that a compact image code is sufficient for visual navigation and that deep networks hold promise for find-ing optimal visual encodings for this task.