Deep Learning for Pose Estimation from RGB-D Images

Abstract

This thesis presents a learning based approach for fast orientation estimation from RGB-D images. For this purpose a special type of a deep convolutional neural network is used - the Autoencoder. As those need lots of labeled data for training two solutions are presented. First, it is shown that they can be trained in an unsupervised fashion on unlabeled real data which then can be assigned pose labels to using synthetic images generated from the object’s CAD model. Secondly, it is shown that using synthetic data for training only is sufficient for orientation estimation. The approach presented has an inference time of less than 2 ms per object and has a low code complexity consisting of only few convolutional and pooling layer.