CarfRCNN: A Two-stage Effective Model For Instance Segmentation
Yujie Chen, Pengbiao Zhao, Julin Chen · Journal of Physics Conference Series · 2021
Abstract With the rapid development of deep learning, many instance segmentation models have achieved good results in accuracy and time. But here are still many problems. In this paper, we proposed a two-stage model CarfRCNN. We proposed CAResNet to change the structure of the backbone, making the feature extraction of the input image more refined. At the same time, we also added the CRF module to add smooth constraints to the pixels, so that the segmentation mask fits the target contour more closely. We train and test on the COCO datasets, and compare with the Mask R-CNN. The experimental results prove that the model we proposed can greatly improve the accuracy.