Bridging the Reality-Gap: 6-DoF Pose Estimation of Multiple Cars by a Deep CNN Trained on Synthetic Data

Kurzfassung

Precise localization in the form of 6-DoF pose estimation is of great value in traffic scenario research. Convolutional neural networks yield robust, fast, and precise results but need huge amounts of annotated data. Obtaining such data in the real world is tedious and error-prone. In contrast, simulation allows automation of the data generation process by creating unlimited amounts of fully and correctly annotated synthetic data. But due to the lack of realism, networks trained on this data usually do not perform well on real data. Recent works bridged this so-called reality gap with synthetic data of high diversity, utilizing domain randomization and photorealistic synthetic data. The Deep Object Pose Estimation (DOPE) shows convincing results in detecting multiple different household objects for the task of robotic grasping. In the thesis I demonstrate that DOPE can be used for the domain of traffic monitoring, in which multiple objects of the same class, namely cars in traffic, are involved. For this purpose I create the synthetic 6-DoF pose estimation dataset MultiCarPose containing highly diverse scenes and utilize it for training DOPE. This allows me to show, that the domain change from robotic grasping to traffic monitoring is possible, that the trained model is able to perform the transition from synthetic to real data and that the pose of multiple objects of the same class in one image can be estimated.