On the Applicability of Unsupervised Feature Learning for Object Recognition in RGB-D Data

We present a feature extraction method for RGB-D data based on k-means clustering that builds on recent work by Coates et al. Using unsupervised learning methods we are able to automatically learn feature responses that combine all available information (color and depth) into one, concise representation. We show that depth information can substantially increase the recognition performance and that the learned features are competitive with previously published results on the RGB-D Object Dataset. For object recognition in high resolution RGB-D images we propose to embed the learned features into a local image descriptor, the convolutional k-means descriptor. Using this approach recognition accuracy can be increased even further.