2020/11/23 by David Biertimpel, Biertimpel, David, Sindi Shkodrani +5
Computer Science · #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Image and Object Detection Techniques
paper · pdf · doi:10.48550/arxiv.2011.11787
openalex publication_date 2020/11/23 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Instance segmentation methods require large datasets with expensive and thus\nlimited instance-level mask labels. Partially supervised instance segmentation\naims to improve mask prediction with limited mask labels by utilizing the more\nabundant weak box labels. In this work, we show that a class agnostic mask\nhead, commonly used in partially supervised instance segmentation, has\ndifficulties learning a general concept of foreground for the weakly annotated\nclasses using box supervision only. To resolve this problem we introduce an\nobject mask prior (OMP) that provides the mask head with the general concept of\nforeground implicitly learned by the box classification head under the\nsupervision of all classes. This helps the class agnostic mask head to focus on\nthe primary object in a region of interest (RoI) and improves generalization to\nthe weakly annotated classes. We test our approach on the COCO dataset using\ndifferent splits of strongly and weakly supervised classes. Our approach\nsignificantly improves over the Mask R-CNN baseline and obtains competitive\nperformance with the state-of-the-art, while offering a much simpler\narchitecture.\n