Comparing transfer-learning and self-supervised learning for ocean floor image classification

Abstract

Benthic habitat mapping is a process of labeling substrates, plants, and animals on the seafloor. Mapping of the benthic habitat is crucial to monitor changes happening due to natural and human-related activities. Annotation of the large amount of data produced with underwater camera systems requires automation. A large dataset of around ten million ocean floor images (BenthicNet) was recently compiled as a part of the BEcoME Project (Benthic Ecosystem Mapping \& Engagement). This thesis discusses the development and specific challenges of a classification system for this dataset. We specifically discuss the importance of careful training and test set partitioning. We further evaluate the performance of pretrained models on ImageNet by supervised learning versus those by self-supervised learning. We show that transfer learning from ImageNet enables good performance comparable with versions that start from self-supervised representations from the BenthicNet dataset.