MCLSS: Multi-level Contrastive Learning for Semantic Segmentation
Di Wang · 2023
Semantic segmentation is a base mission of computer vision, with the aim of dividing the input picture into different parts which represent a semantically interpretable category. Besides, semantic interpretability means that the part is meaningful in the real world. Contrastive Learning is a desirable method, so we introduce it to our work to improve performance, which is meaningful. In this paper, we propose Multi-level Contrastive Learning for Semantic Segmentation (MCLSS) method to solve the problem. Specifically, we regard the images with the same label as positive pair and the image with a different label as negative pair, this paper forces on narrowing the gap between identical examples. Experiments on PASCAL VOC2012 dataset demonstrate that our model outperforms the original model, which means our model efficiently narrows the gap between the images with the same label. For the analysis, our model can effectively improve the semantic segmentation performance and achieve an average mIOU of 80.2.