Figure 2.

Overview of the model used in this study. Both semantic and instance segmentation branches use the same feature map and share the encoder. Each branch has its multiscale context module and decoder, both of which use ASPP. The instance segmentation branch predicts the center of mass for each instance and the offset vector between the center and each pixel. With the predicted center and the offset, the instance ID of each pixel can be determined, thereby resulting in class-agnostic instance segmentation. The final panoptic segmentation is obtained by fusing the results of the semantic segmentation and class-agnostic instance segmentation. ASPP, atrous spatial pyramid pooling.