 Computer Vision and Geometry GroupOpen OpportunitiesAtracsys is a Swiss company based in Puidoux (VD) developing high-end optical tracking systems that can easily be customized and integrated into various applications in the medical, industrial and research fields. The master thesis goal is to search for new camera models and alternative optimization methods to improve the calibration of new kinds of stereo cameras developed at Atracsys. - Computer Vision
- Master Thesis
| In this project, we propose to investigate if scene graph technology can assist place recognition in changing environments by considering the semantic information about the environment. - Computer Vision, Intelligent Robotics
- Master Thesis, Semester Project
| Atracsys’ SpryTrack 300 is a hybrid stereo camera system that enables multimodal image acquisitions in one package: Marker-based localization frames for standard Computer Assisted Orthopedic Surgery, RGB-D frames with the help of an integrated random pattern projector, and stereo RGB imaging. The purpose of the master thesis is to develop and compare competing 3D face scanning algorithms with a hardware setup based on spryTrack 300 cameras. The student may be able to try and test standard photogrammetry 3D scanning methods, one-shot texture and 3D mesh generation algorithms, or NeRF-based methods. - Computer Vision, Pattern Recognition
- ETH Zurich (ETHZ), Master Thesis
| Atracsys spryTrack 300 is a hybrid stereo camera system that enables multimodal optical 3D tracking in one package: In the near infrared mode, it captures low exposure stereo frames to provide standard intraoperative marker-based 3D localization data for standard Computer Assisted Orthopedic Surgery. In the RGB mode, it captures long exposure RGB stereo frames illuminated with an integrated random pattern dot projector.
In the context of Computer Assisted Orthopedic Surgery, the purpose of the master thesis is to improve the intraoperative markerless reconstruction of bones or instruments provided by the spryTrack RGB mode though a deep learning method whose training ground truth data is given by the marker-based 3D poses obtained with the same spryTrack camera. - Computer Vision, Pattern Recognition
- ETH Zurich (ETHZ), Master Thesis
| This project explores the integration of the IDUN Guardian EEG sensor into an AR surgical simulator to adaptively assess and respond to trainees' stress and cognitive loads, aiming to personalize training and enhance skill acquisition. By analyzing EEG data with advanced signal processing and machine learning, the project seeks to improve surgical training outcomes and standardize education across the field.
Background
Surgeries consist of a series of complex tasks and external factors such as unexpected complications or changes in the patient's condition can impact the course of the surgery. The digitization of surgical training can reduce the risk of surgical errors, improve patient outcomes, and provide standardized training and assessment. This transformation allows for safer and more efficient training of surgeons, ensuring that the surgical staff receives consistent education allowing for easier comparison of skills across different individuals and institutions. At the ROCS group, located at Balgrist University Hospital in Zurich, we are developing a surgical training simulator for adaptive Augmented Reality-based training for orthopedic surgeries. An automated assessment of stress and cognitive load of the trainee can play an important role in automated assessment and adaptive, personalized surgical training. In this context, we want to explore the capabilities of the IDUN guardian, an in-ear EEG sensor developed by an ETH Spin-off (https://iduntechnologies.com/idun-guardian-in-ear-eeg-platform/).
- Computer Vision, Pattern Recognition, Signal Processing, Virtual Reality and Related Simulation
- Semester Project
| This master's thesis delves into the surgical application of Neural Radiance Fields (NeRF) and Gaussian splatting to enhance 3D reconstruction of lumbar vertebrae using intraoperative X-rays. The study focuses on understanding NeRF fundamentals, comparing NeRF-based approaches with traditional methods, and optimizing NeRF for precise 3D reconstruction. The expected outcomes comprise a comprehensive grasp of NeRF and insights into its potential for surgical applications. The findings are intended to benefit orthopedic surgical planning, contributing to improved surgical outcomes - Artificial Intelligence and Signal and Image Processing, Computer Software, Engineering and Technology, Orthopaedics
- Master Thesis, Semester Project
| In continual learning, deep learning models incrementally learn more classes or tasks over time. Doing so, they should not forget previously learned knowledge. This is a hard and active research problem. Making it even harder, we want the models to also estimate correct uncertainty. E.g., they should be highly uncertain about a new object type, but not uncertain about an object that they just learned correctly.
[1] Parisi et al., Lifelong learning with neural networks http://dx.doi.org/10.1016/j.neunet.2019.01.012
[2] Gawlikowski et al., Uncertainty in Deep Neural Networks http://arxiv.org/abs/2107.03342 - Engineering and Technology, Information, Computing and Communication Sciences
- Master Thesis, Semester Project
| Generally, for SfM systems, camera models (eg. Pinhole, Fisheye) and their intrinsics are assumed known. However, 1D Radial SfM establishes a SfM pipeline for scene reconstruction with camera pose estimated up to translation, while implicit camera model proposes an implicit camera distortion model which can estimate the remaining degree-of-freedom and performs bundle adjustment without prior knowledge on the camera distortion model.
Base on the result, it should be possible to do iterative triangulation and establish a SfM pipeline without knowledge of camera model
- Computer Vision
- Semester Project
| The goal of this project is to develop an end-to-end differentiable global structure-from-motion algorithm. - Computer Vision
- Master Thesis, Semester Project
| The goal is the develop an algorithm that builds a pose graph from an unordered set of images as fast as possible. This is achieved by, first, building a minimal spanning tree from the images exploiting predicted similarity scores. Then the spanning tree is populated with additional edges until the pose uncertainty falls below a threshold in each vertex. This procedure is very important for Structure-from-Motion algorithms where the first step is generated such pose graphs. - Computer Vision
- Master Thesis, Semester Project
|
|
