Weakly Supervised One Shot Segmentation

Hasnain Raza, Mahdyar Ravanbakhsh, Tassilo Klein, Moin Nabi; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 0-0

One-shot learning is a challenging discipline of machine learning since it gnaws at the concept of learning from large amounts of data. This is akin to making machine learning algorithms generalize from a few examples, much like how humans learn. We explore another novel dimension to this problem, of using weak supervision (labels only) in the one-shot domain, and specifically analyse it in the context of semantic segmentation. This is a challenging problem since we operate in the scarcity of data and supervision. We present a simple yet effective approach, whereby exploiting information from the base training classes in the current one-shot segmentation set-up allows for weak supervision to be easily used. We show that this strategy can be leveraged to achieve nearly the same results as full supervision, but with no pixel annotations, allowing fully automated segmentation. Comparisons to several fully supervised methods show convincing results. As well as better results than a weakly supervised baseline. Also presented is a baseline for generalized segmentation under one-shot and weak supervision assumptions.