Underwater Object Recognition: A Domain-Adaption Methodology of Machine Learning Classifiers

This paper presents a novel dataset, composed of images of objects in two distinct environments and both controlled and uncontrolled capture conditions, aimed at serving as a benchmark for domain-adaptation image classification algorithms in an air versus underwater context. All images are fully annotated, extending the use of the dataset for detection as well as segmentation tasks. An exemplifying use-case is tested, where the performance of a Support Vector Machine applied to a Bag-of-Visual-Words and SIFT features is evaluated on both domains, with different training methodologies. Results demonstrate that the conventional classifier used has a lack of generalization ability, with a poor transfer of knowledge from the aerial to the aquatic domain.