Weakly Supervised Instance Segmentation by Deep Community Learning

We present a weakly supervised instance segmentation algorithm based on deepcommunity learning with multiple tasks. This task is formulated as acombination of weakly supervised object detection and semantic segmentation,where individual objects of the same class are identified and segmentedseparately. We address this problem by designing a unified deep neural networkarchitecture, which has a positive feedback loop of object detection withbounding box regression, instance mask generation, instance segmentation, andfeature extraction. Each component of the network makes active interactionswith others to improve accuracy, and the end-to-end trainability of our modelmakes our results more robust and reproducible. The proposed algorithm achievesstate-of-the-art performance in the weakly supervised setting without anyadditional training such as Fast R-CNN and Mask R-CNN on the standard benchmarkdataset. The implementation of our algorithm is available on the projectwebpage: https://cv.snu.ac.kr/research/WSIS_CL.