A Unified Instance Segmentation Framework for Completely Occluded Objects and Dense Objects in Robot Vision Measurement

Instance segmentation for completely occluded objects and dense objects in robot vision measurement are two challenging tasks. To uniformly deal with them, this paper proposes a unified coarse-to-fine instance segmentation framework, CFNet, which uses box prompt-based segmentation foundation models (BSMs), e.g., Segment Anything Model. Specifically, CFNet first detects oriented bounding boxes (OBBs) to distinguish instances and provide coarse localization information. Then, it predicts OBB prompt-related masks for fine segmentation. CFNet performs instance segmentation with OBBs that only contain partial object boundaries on occluders to predict occluded object instances, which overcomes the difficulty of existing amodal instance segmentation methods in directly predicting occluded objects. In addition, since OBBs only serve as prompts, CFNet alleviates the over-dependence on bounding box detection performance of current instance segmentation methods using OBBs for dense objects. Moreover, to enable BSMs to handle OBB prompts, we propose a novel OBB prompt encoder. To make CFNet more lightweight, we perform knowledge distillation on it and introduce a Gaussian label smoothing method for teacher model outputs. Experiments demonstrate that CFNet outperforms current instance segmentation methods on both industrial and public datasets. The code is available at https://github.com/zhen6618/OBBInstanceSegmentation.