A Study On Instance Segmentation Based On Image-level Supervision

To reduce the dependence of fully-supervised instance segmentation on pixel-level annotations,weakly-supervised instance segmentation has become an important research topic.IRNet(Inter-pixel Relation Network)is an advanced weakly-supervised instance segmentation framework using only image-level labels.Based on a careful study on IRNet,several effective strategies have been proposed to achieve better instance segmentation results.The main contributions are as follows:(1)A multi-scale instance center detection network is developed to improve the accuracy of the predicted displacement field.Since the normalized displacement vectors with respect to the instance centers should be invariant for the objects with different scales,a siamese neural network and a self-supervised learning loss function have been proposed to better identify the instance centers.The results on PASCAL VOC 2012 dataset show that more robust displacement fields have been obtained and the instance segmentation precision has been increased from 36.4 to 38.5.(2)An boundary detection network is developed to obtain the instance-wise boundaries and reduce over-segmentation.In IRNet,only class-wise boundaries can be obtained based on Class Activation Maps.To detect instance-wise boundaries,inter-pixel relations in the predicted displacement field have been exploited in this thesis.The results show that more complete instance-wise boundaries have been obtained and the instance segmentation precision has been increased from 38.5 to 40.6.(3)Prior knowledge about human shapes are incorporated to improve the developed instance segmentation method.Based on the instance centers and instance-wise boundaries,prior knowledge about human shapes have been integrated into the developed method by the template matching strategy.With the prior knowledge,the instance segmentation results have been increased from 23.8 to 26.6.To achieve better instance segmentation results based on image-level annotations,three effective strategies to improve the instance center detection,to complete the instance-wise boundaries and to incorporate conceptual prior knowledge have been developed in this thesis.The proposed methods will be useful for the future application of the weakly-supervised instance segmentation task.