Model assisted bootstrapping for annotation of segmentation datasets

Applying deep learning to new domains usually implies a considerable data collection problem. We look at idea of how we can use a partially trained model as an aid to a human annotator. We do this by providing the partially trained model's prediction as a starting point for a human annotator to directly edit. This is demonstrated by applying our ideas to building a small segmentation dataset for labeling trees in a plantation. We also show that by starting with a pre-trained model and fine-tuning, we can provide a useful aid to a human annotator using very few input images.