Industrial Doctoral School

One project under the IDSL programme, has to do with Vision-based Positioning with Diverse Imaging Sensors. Today, vision-based methods are available which provide position estimates in the archipelago with GPS-level accuracy under rather ideal conditions, e.g., an omnidirectional camera with an instantaneous 360° field of view (FOV) around the vessel, daylight and good weather conditions, training images available from a similar terrain as the operational area.



In a real scenario, the vision-based system has to operate under non-ideal conditions. The available camera sensor on the vessel may have a small FOV requiring image information to be fused over time whilst scanning the camera around the vessel. Low-light conditions or night operations may require the use of infrared image sensors. Neighbouring vessels may occlude parts of the surrounding geographic horizon. Robust ego positioning when operating in previously unvisited areas under these non-ideal conditions require new and more resilient matching techniques to be developed. To this end, the first learning-based framework that explores the fusion of inertial data into an encoder-decoder model and extracts the contours is proposed. Preliminary results have recently been published at 19th ACM SIGGRAPH European Conference on Visual Media Production (CVMP’22)



A second project within the IDSL programme has to do with one-shot learning from textual data. Machine learning, particularly Deep Neural Networks (DNNs), has achieved outstanding outcomes in numerous real-world applications such as image and text classification. However, most of these learning methods require a considerable amount of annotated training data, but this is not always available in some applications. For example, an automated software tool should be able to learn from a small number of interactions with the user to efficiently customise with new user’s requirement.



In this project, we have demonstrated that Inductive Logic Programming (ILP), in particular Meta-Interpretive Learning (MIL), can be used to generate human-readable classification rules from small amount of textual training data. This novel approach is promising for applications where we cannot access a large amount of training data. A paper describing the preliminary results was presented and published at the 39th International Conference on Logic Programming (ICLP 2023).