We explore salient-instance segmentation from novel balance perspectives. Salient-instance segmentation includes salient detection and salient segmentation sub-tasks. We focus on the RoIMasking layer in S4Net, which outputs the training data for salient segmentation. Considering the balance between sub-tasks, we propose an expanded smooth RoIMasking (SM) for salient segmentation. SM smooths the training data of salient segmentation, which can reduce the complexity of segmentation and increase the importance of detection in the whole task, leading to the improvement of salient detection sub-task. Moreover, the existing salient segmentation sub-model cannot recover from the salient detection error. We attribute this to the extreme imbalance distribution of salient segmentation training data. To address this problem, an expanded mutilated RoIMasking (MM) is proposed, which introduces false negative samples into training data and balances the distribution of salient segmentation data. Additionally, due to the lack of public dataset for salient-instance segmentation, we build a new dataset: Instance-MSRA1000. By simply embedding SM and MM into the existing network, a improved method with 4.4 points mAP (mean Average Precision) gains can be achieved compared to S4Net. The experimental results on the existing public and new built datasets show that SM and MM significantly performs well.