Machine learning and computer vision techniques open up interesting possibilities for the future of spacecraft identification, pose estimation and rendezvous. Unfortunately, datasets for space applications are extremely limited or nonexistant. The Aerospace and Computer Science communities at large stand to benefit from high-quality open source datasets. EPFL currently stands at the forefront of hyperrealistic rendering and is thus well situated to provide guidance to the community at large with a multi-object, unseen, occlusion spacecraft dataset.
In this project, the student will construct a number of 3D spacecraft and debris objects, integrate a number tumble profiles, and insert these motions into various earth orbits. The student will then render images of the objects in orbit to create a public dataset. Upon initial completion of the dataset, there are several research thrusts depending on the student’s interest.
- Improving the rendering realism (environmental lighting, in-orbit albedo, multi-object reflections).
- Pose Estimation Architecture Performance Improvement.
- Pose Estimation Network Latency and Size Reduction.
References
- An Adaptive Parameterization for Efficient Material Acquisition and Rendering https://rgl.epfl.ch/publications/Dupuy2018Adaptive [SIGGRAPH Asia 2018]
-
Wide-Depth-Range 6D Object Pose Estimation in Space [CVPR21]
-
A Monocular Pose Estimation Case Study: The Hayabusa2 Minerva-II2 Deployment [CVPR21 Workshop]
Prerequisites
- Proficiency in Python
- Basics of computer vision (pinhole camera model, image projections, etc)
- Interest in image rendering