Institute for Intelligent Interactive Systems

Open Opportunities

Explicit 3D avatars from a video ETH Zurich Advanced Interactive Technologies The learning of animatable 3D body avatars has diverse applications in gaming, video production, and AR/VR communication. While recent methods using neural implicit representations, such as Signed Distance Functions (SDFs), can capture high-quality geometry, they are often inefficient to train as well as challenging to animate. Additionally, converting these implicit avatars into meshes is necessary for rendering them in standard engines, leading to a reduction in rendering quality. Recent work has made great progress in using explicit representations, e.g. point cloud and meshes, to learn 3D geometry. Nvdiffrec employs meshes to learn high-quality static geometry. PointAvatar leverages animatable point clouds to represent head avatars. Can we extend them to full body, clothed avatars? After reconstructing the avatar, there are also several exciting follow-up tasks. For example, can we modify the avatar given text guidance? Can we learn avatars from only a few images instead of a video? Computer Graphics, Computer Vision CLS Student Project (MPG ETH CLS), IDEA League Student Grant (IDL), Master Thesis, Semester Project

Computational simulation framework of robotic and human motion for visual and inertial modeling ETH Zurich Sensing, Interaction & Perception Lab Motion modeling, for robots and humans alike, has become a highly active field of research and developed methods have found many applications in state estimation, navigation, localization, self-driving cars, drones and autonomous aerial vehicles, activity detection, pose estimation, or even rehabilitation. Sensors often include cameras as well as inertial motion units, such as accelerometers and gyroscopes, which are now commonly embedded in nearly each device today, from heavy machinery in industrial settings to personal devices, such as tables, phones, and watches. However, the lack of large-scale, labeled datasets impedes progress in developing robust and generalized predictive models for neural inertial-based motion estimation. Especially, labeled data in human activity data is scarce and hard to come by, as sensor data collection is expensive, and the annotation is time-consuming and error-prone. Aerodynamics, Computer Vision, Flight Control Systems, Flight Dynamics, Intelligent Robotics, Modeling and Simulation, Robotics and Mechatronics, Simulation and Modelling, Virtual Reality and Related Simulation ETH Zurich (ETHZ), Master Thesis, Semester Project