This paper focuses on estimating probability distributions over the set of 3D ro- tations (SO(3)) using deep neural networks. Learning to regress models to the set of rotations is inherently difficult due to differences in topology between R^N and SO(3). We overcome this issue by using a neural network to out- put the parameters for a matrix Fisher distribution since these parameters are homeomorphic to R^9 . By using a negative log likelihood loss for this distri- bution we get a loss which is convex with respect to the network outputs. By optimizing this loss we improve state-of-the-art on several challenging applica- ble datasets, namely Pascal3D+, ModelNet10-SO(3). Our code is available at https://github.com/Davmo049/Publicproborientationestimationwith_matrix fisherdistributions

In our current research we examine the application of visuo-haptic augmented reality setups in medical training. To this end, highly accurate calibration, system stability, and low latency are indispensable prerequisites. These are necessary to maintain user immersion and avoid breaks in presence which potentially diminish the training outcome. In this paper we describe the developed calibration methods for visuo-haptic integration, the hybrid tracking technique for stable alignment of the augmentation, and the distributed framework ensuring low latency and component synchronization. Finally, we outline an early prototype system based on the multimodal augmented reality framework. The latter allows colocated visuo-haptic interaction with real and virtual scene components in a simplified open surgery setting.

The driving force of our research is the precise combination of real and - possibly indistinguishable - virtual objects in an interactive augmented reality environment. This requires real-time, multimodal simulation, as well as stable and accurate overlay of the computer-generated objects. This paper describes several methods to improve accuracy and stability of our hybrid augmented reality system. In a comparison of two approaches to hybrid head pose refinement, we show that Quasi-Newton method enables high performance optimization for image space error minimization. Moreover, a 3D landmark refinement step is proposed, which significantly improves quality and robustness of the overlay process. The enhanced system is demonstrated in an interactive AR environment, which provides accurate haptic feedback from real and virtual deformable objects. Finally, the effect of landmark occlusion on tracking stability during user interaction is also analyzed.

An integral part of every augmented reality system is the calibration between camera and camera-mounted tracking markers. Accuracy and robustness of the AR overlay process is greatly influenced by the quality of this step. In order to meet the very high precision requirements of medical skill training applications, we have set up a calibration environment based on direct sensing of LED markers. A simulation framework has been developed to predict and study the achievable accuracy of the backprojection needed for the scene augmentation process. We demonstrate that the simulation is in good agreement with experimental results. Even if a slight improvement of the precision has been observed compared to well-known hand-eye calibration methods, the subpixel accuracy required by our application cannot be achieved even when using commercial tracking systems providing marker positions within very low error limits.