Bayesian Semantic Instance Segmentation in Open Set World
share
This paper addresses the instance segmentation task in the open-set conditions, where input images might contain known and unknown object classes. Existing instance segmentation methods require annotating masks for all object instances, which is expensive and even infeasible in realistic scenarios, where the number of categories increases boundlessly. In this paper, we present a novel open-set instance segmentation approach to segment all object instances in images. Based on the output of an object detector/segmenter trained on known classes, we seek a global image segmentation, whose regions are each assigned to one of the known or unknown classes. We formulate the problem in a Bayesian framework, and approximate the posterior distribution using a simulated annealing optimization equipped with an efficient image partition sampler. We show empirically that our method is competitive with state-of-the-art supervised methods on known classes, but also performs well on unknown classes when compared with unsupervised methods.
READ FULL TEXT
