We consider the competition between instance and semantic segmentation in panoptic segmentation to develop the deep chain instance segmentation network (ChaInNet) to mitigate this problem. Segmentation competition is caused by the usual contradiction between instance and semantic segmentation when predicting instance objects. ChaInNet alternately performs inter-reference learning by stacking two-branch chain blocks to improve feature extraction from network layers. Panoptic segmentation using ChaInNet accurately extracts the contour of instance objects and improves the accuracy of instance segmentation, thus reducing the adverse effects of segmentation competition on the quality of the outcome. ChaInNet is a general instance segmentation architecture that can be widely used in various object recognition tasks. Experimental results on the MS COCO and Cityscapes benchmark datasets show that ChaInNet provides state-of-the-art segmentation and outperforms Mask R-CNN, which is commonly used for identifying instance objects in panoptic segmentation.

我们考虑全景分割中实例和语义分割之间的竞争，以开发深链实例分割网络（ChaInNet）来缓解这个问题。分割竞争是由在预测实例对象时通常的实例和语义分割之间的矛盾引起的。ChaInNet 通过堆叠两个分支链块交替执行参考间学习，以改进网络层的特征提取。使用ChaInNet的全景分割准确地提取了实例对象的轮廓，提高了实例分割的准确性，从而减少了分割竞争对结果质量的不利影响。ChaInNet 是一种通用的实例分割架构，可广泛用于各种对象识别任务。