Clustering for large-scale segmentation dataset collection

Abstract

Segmentation datasets are smaller and much more expensive to collect than their image classification counterparts. Leveraging machine learning in the annotation process will be critical to scaling these datasets up. In this thesis, we propose an iterative cluster-based approach to segmentation data collection. By using existing networks to predict millions of segmentations and clustering to group similar predictions together, we ask human annotators a small number of questions per cluster and collect a large number of reasonable-quality segmentations at low cost. Although the collected segmentations are biased towards objects already predicted by the network, we demonstrate that they improve performance upon re-training and that the procedure can be applied iteratively, up to a point, to discover harder and harder objects. We demonstrate this pipeline in simulation and show promising results on real unlabeled images. We also present a new annotation tool called LabelMeLite for the rapid filtering and editing of predicted segmentations.