| Popis: |
Instance segmentation for autonomous driving should have a simple process. For dense instance segmentation methods based on FCOS, although they can directly predict instance masks, they output many duplicate predictions. Therefore, they must use non-maximum suppression (NMS) to remove the duplicate predictions and improve the performance. These methods with NMS not only hinder end-to-end training but also make the whole process more complicated. In contrast to these works, we propose a NMS-free FCOS based instance segmentation (NFIS). NFIS achieves comparable performance with NMS. Besides, we propose a context-based encoding module (CBEM) to refine the feature maps. CBEM utilizes the local and global information to refine the parameters of dynamic convolution and improve the classification. The proposed network outperforms most instance segmentation networks on the street scene datasets Cityscapes and Indian Driving Dataset (IDD). |
