UPerNet with ConvNeXt for Semantic Segmentation
Runze Wang, Haoyu Jiang, Yufei Li
Abstract
In this article, we adopt UPerNet as the model and ConvNeXt as the backbone for semantic segmentation of cityscapes. UPerNet is a highly compatible architecture that can handle a variety of visual tasks. At the same time, as a convolutional network without adding attention mechanism, convnext has excellent feature extraction ability.In this article, we combine UPerNet and ConvNeXt to obtain a semantic segmentation model. Their excellent feature extraction ability and multi-scale object detection capability are utilized to enhance the performance of the network. In addition, we select the appropriate ConvNeXt structure, i.e. ConvNeXt-small, according to the size and characteristics of cityscapes dataset, and optimized the loss function to solve the problem of uneven class distribution. The experimental results show that ConvNeXt has excellent performance in feature extraction. Compared with the classic backbone ResNet50, it has improved by 1.96% on aAcc, 12.28% on mIoU and 11.39% on mAcc. After the loss function is optimized, all metrics also increase, which fully proves that our method is effective.