An Improved Shared Encoder-Based Model for Fast Panoptic Segmentation

Panoptic segmentation aims to generate coherent scene segmentation by unifying semantic segmentation and instance segmentation, which is one of the important tasks in the scene perception and understanding for the unmanned autonomous system. Deep learning methods have been widely used in the panoptic segmentation task and achieved great research progress. However, how to balance the efficiency and accuracy of the deep learning-based model is still a challenging problem in the field of panoptic segmentation. To deal with this problem, an improved shared encoder-based model for fast panoptic segmentation is proposed in this article. In the proposed model, a simple and effective shared encoder is proposed to provide semantically rich multiscale features. In addition, a lightweight semantic branch is presented to fully exploit the multiscale features to extract semantic information, and a dynamic kernel-based instance branch is proposed to extract instance information. At last, a panoptic fusion head is proposed to fuse the information of the two branches based on the occlusion-aware discrimination mechanism to generate the final panoptic segmentation output. Finally, the proposed model is tested on two public datasets Cityscapes and COCO. The results show that the proposed model can achieve a good balance between efficiency and accuracy.