||Johanna Engman <firstname.lastname@example.org>|
||2023-09-12 – 2024-04-01|
In SLAM systems a key part is feature extraction, which can be done in various ways, usually by using point clouds of some sort. Instead of require a large amount in data per frame, such as point clouds do, we want to use more describing features, geometries. In this project, as a part of a bigger SLAM system we aim to do feature extraction by detecting lines in images containing cylinders, typically images from forests, using machine learning.
These images are often poorly annotated and therefore we wish to gain good results in our feature extraction using only a small dataset with annotated images and a larger dataset without any annotations. We are combining the latest architecture within semi-supervised learning for semantic segmentation, based on the ideas for contrastive learning, together with a very efficient line detector, building a model for the feature extraction.