Revisiting Deep Convolutional Neural Networks for RGB-D Based Object Recognition

In this paper we reinvestigate Deep Convolutional Neural Networks (DCNNs) for RGB-D based object recognition. A previously proposed method in which DCNNs are pretrained on a large-scale RGB database and just fine-tuned to process colorized depth images is taken up and extended. We introduce and analyse multiple solutions to improve depth colorization and propose a new method for depth colorization based on surface normals. We show that our improvements increase the classification accuracy significantly, such that we can present new state-of-the-art results for the Washington RGB-D dataset. Our results also indicate that classification using only surface normals without RGB images outperforms classification using pure RGB images, which is to our knowledge a novel discovery in the field of DCNNs.