Coral-Segmentation: Training Dense Labeling Models with Sparse Ground Truth

摘要

Biological datasets, such as our case of study, coral segmentation, often present scarce and sparse annotated image labels. Transfer learning techniques allow us to adapt existing deep learning models to new domains, even with small amounts of training data. Therefore, one of the main challenges to train dense segmentation models is to obtain the required dense labeled training data. This work presents a novel pipeline to address this pitfall and demonstrates the advantages of applying it to coral imagery segmentation. We fine tune state-of-the-art encoder-decoder CNN models for semantic segmentation thanks to a new proposed augmented labeling strategy. Our experiments run on a recent coral dataset [4], proving that this augmented ground truth allows us to effectively learn coral segmentation, as well as provide a relevant score of the segmentation quality based on it. Our approach provides a segmentation of comparable or better quality than the baseline presented with the dataset and a more flexible end-to-end pipeline.