Multi-Stage Instance Segmentation
Guohua Zhu, Suk Chan Kim · 2022 13th International Conference on Information and Communication Technology Convergence (ICTC) · 2022
Two-stage instance segmentation relies on the correctness of the Region Proposal Network (RPN). To improve the accuracy of RPN, we propose a slight modification of the two-stage object detection network Faster R-CNN as a region proposal network. First, we remove the class branch of the second stage of Faster R-CNN and use the box regression as the output of the entire proposal network. Then we add a new box regression branch, classification branch, and segmentation branch as the final output. Finally, we train the entire network. The results show a great improvement in segmentation accuracy.