OccludedInst: An Efficient Instance Segmentation Network for Automatic Driving Occlusion Scenes

Instance segmentation is widely used in autonomous driving as a basic computer vision task, because it can perform both instance-level distinction and pixel-level segmentation. In this paper, we propose a query-based instance segmentation algorithm OccludedInst based on QueryInst for solving the occlusion issues more comprehensively. OccludedInst is designed from two parts: data preprocessing and network structure. In the data preprocessing, the Simple Copy-Paste data augmentation method is used to increase the number of occlusion situations and enable the network to acquire additional occlusion processing experience and countermeasures. In the network structure, we take the form of occluded features as the boundary and divide it into two cases. For the occluded features to be provided in the form of the FPN feature layers, the occlusion correction module is developed to correct the original FPN feature layers. For the occluded features to be presented as Region of Interest (RoI) features, the extended dynamic head is proposed to improve the acquisition of channel and spatial dimension instance information of occluded RoI features. In addition, we also design the feature transmit structure that strengthens the link of occluded RoI features for detection and segmentation branches. Finally, we propose Cityscapes-OCC and BDD100K-OCC subsets and use KINS datasets to verify the effectiveness of OccludedInst for occlusion scenes. It has increased by 2.5%, 2.2% and 2.4%, respectively, compared to the baseline. Moreover, our top model is ranked eleventh on the Cityscapes ranking list.