There are many challenges related to the detection of small objects as they encompass multiple clutter distributions, imbalanced classes, as well as the issue of noisy labels and low resolution.

Moreover, it is of paramount relevance - especially for operational scenarios - to provide a quantification of the confidence that can be associated with the outcomes produced by object detection platforms. For these reasons, research on new deep learning approaches to address these points is necessary.

Specifically, the results of these schemes will help in detecting small objects in water areas, as well as tracking vessels from harbor to open water, possibly combining multiple sensors (e.g., remote sensing data and AIS records) to improve the precision of the methods.

For vessel and object detection from satellites, Visual Intelligence will exploit available labelled training data that have been collected over time by KSAT operators through the manual quality assurance procedures in their semi-automated ship detection services. Our research activities will target the training of neural networks under the constraint of limited training data. It is of interest to be able to assess the confidence and the reliability of the detections.