This project aims to develop a (soft) scene-aware point cloud correspondence ejector. The main focus will lie on the training procedure (self-supervised, weakly-supervised,...) of the weighting network, that can then be deployed in classic or learned point cloud registration algorithms.
Reliable point cloud registration is among the core components of any robotic system. These registration algorithms usually rely on point cloud correspondences that are incorporated into a large optimization problem to obtain displacement and relative rotation [1].
To date, the state-of-the-art solutions weigh all correspondence pairs equally, and the core criterion for identifying correspondences and assessing their quality is the euclidean distance in space. While a low euclidean distance is often a necessary condition, it does not guarantee the addition of meaning- and helpful constraints into the optimization. For example, point correspondences in the crown of a tree provide noisy information due to their elasticity and external factors such as wind or yearly seasons.
Following this line of thought, the goal of this project is to develop a (soft) scene-aware point cloud correspondence rejector. Here the correspondences should be judged based on their component’s appearance and the pair’s distribution. The resulting mask - provided by a (point-wise) neural network classifier - should then be used to either exclude correspondence pairs from the optimization or to weigh them according to their expected importance.
The main focus of this work will lie on the training procedure (self-supervised, weakly-supervised,...) of the weighting network, that can then be deployed in classic or learned [2] point cloud registration algorithms.
Reliable point cloud registration is among the core components of any robotic system. These registration algorithms usually rely on point cloud correspondences that are incorporated into a large optimization problem to obtain displacement and relative rotation [1]. To date, the state-of-the-art solutions weigh all correspondence pairs equally, and the core criterion for identifying correspondences and assessing their quality is the euclidean distance in space. While a low euclidean distance is often a necessary condition, it does not guarantee the addition of meaning- and helpful constraints into the optimization. For example, point correspondences in the crown of a tree provide noisy information due to their elasticity and external factors such as wind or yearly seasons. Following this line of thought, the goal of this project is to develop a (soft) scene-aware point cloud correspondence rejector. Here the correspondences should be judged based on their component’s appearance and the pair’s distribution. The resulting mask - provided by a (point-wise) neural network classifier - should then be used to either exclude correspondence pairs from the optimization or to weigh them according to their expected importance. The main focus of this work will lie on the training procedure (self-supervised, weakly-supervised,...) of the weighting network, that can then be deployed in classic or learned [2] point cloud registration algorithms.
- Literature research and problem investigation.
- Implementation of a weakly supervised baseline.
- Development of a more tightly integrated self-supervised correspondence outlier detector.
- Literature research and problem investigation. - Implementation of a weakly supervised baseline. - Development of a more tightly integrated self-supervised correspondence outlier detector.
- Experience in Python.
- Experience in C++ and ROS is a plus
- Knowledge with LiDARs and RGB cameras.
- Highly motivated and research-oriented.
- Experience in Python. - Experience in C++ and ROS is a plus - Knowledge with LiDARs and RGB cameras. - Highly motivated and research-oriented.