Augmented Autoencoders for Object Orientation Estimation trained on synthetic RGB Images

Kurzfassung

Fast and accurate object pose estimation algorithms are crucial for robotic tasks. Despite intensive research, most approaches are not generally applicable on arbitrary object characteristics and dynamic environment conditions. Learning-based methods like Convolutional Neural Networks (CNNs) have proven good generalization properties given sufficient training data. However, annotating RGB images with 3D object orientations is difficult and requires expert knowledge. In this work, a real-time approach for joint 2D object detection and 3D orientation estimation is proposed. First, a CNN-based object detector [45] is used to localize objects in an image plane. In the second step, an Autoencoder (AE) predicts the 3D orientation of the object from the resulting scene crop. The main contribution is a new training method for AEs that allows learning 3D object orientations from synthetic views of a 3D model, dispensing with the need to annotate orientations in real sensor data. The AE is trained to revert augmentations applied to the input and thus becomes robust against irrelevant color changes, background clutter and occlusions. It learns to produce low-dimensional representations of synthetic object orientations which can be compared to the representations of real RGB test data in a k-Nearest-Neighbor (kNN) search. Experiments on the pose annotated dataset T-LESS [23] prove the performance of the approach on different sensors. Finally, the training on synthetic data is shown to be almost on par with the training on real data.