Overview
Standard benchmarks in image segmentation assume a "closed-world" setting, in which a pre-determined set of non-overlapping object categories is exhaustively segmented and labeled in all training and evaluation images. This significantly increases the difficulty of data collection, requiring either complex quality control and post-processing schemes if using crowd-sourced labeling or superhuman feats of conscientiousness and endurance by a single expert annotator [1]. Data collection could be made much easier if the requirements for exhaustive segmentation and consistent categorization were relaxed, but standard segmentation models struggle when labels and exhaustive annotations are removed [2]. We explore the use of self-supervised learning to augment the training of image segmentation in order to improve the ability of models to generalize to new categories not annotated during training.
References
[1]: A. Barriuso and A. Torralba. (2012). Notes on image annotation
[2]: D. Kim, T. Lin, A. Angelova, I. S. Kweon, and W. Kuo. (2021). Learning Open-World Object Proposals without Learning to Classify
Researchers
Norman Mu (UC Berkeley)
Saining Xie (Meta AI - FAIR)
Alexander Kirillov (Meta AI - FAIR)
David Wagner (UC Berkeley)
Trevor Darrell (UC Berkeley)