# Semi-Supervised Semantic Segmentation with Cross Pseudo Supervision

**Authors:** Xiaokang Chen, Yuhui Yuan, Gang Zeng, Jingdong Wang  
**arXiv:** [2106.01226](https://arxiv.org/abs/2106.01226)  
**Date:** 2021-06-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.01226v1/

## Key points

- Problem: Semi-supervised semantic segmentation
- Approach: Cross Pseudo Supervision (CPS) consistency regularization
- Imposing consistency on two perturbed segmentation networks
- Using pseudo one-hot label map to supervise each network
- Encouraging similarity between predictions of perturbed networks
- Expanding training data by incorporating unlabeled data with pseudo labels
- State-of-the-art performance on Cityscapes and PASCAL VOC 2012 datasets
- Effectiveness against fully-supervised baselines, even with large amount of labeled data available (shown in Figure 3)
- Ablation study evaluating different loss combinations on PASCAL VOC 2012 and Cityscapes datasets
- Consistently outperforms other methods under various partition protocols
- Promising solution for improving semi-supervised semantic segmentation by leveraging both labeled and unlabeled data through cross pseudo supervision

## Summary

This paper addresses the problem of semi-supervised semantic segmentation by utilizing both labeled and unlabeled data. The authors propose a novel consistency regularization approach called Cross Pseudo Supervision (CPS). CPS involves imposing consistency on two segmentation networks that are perturbed with different initializations for the same input image. One network's output, in the form of a pseudo one-hot label map, is used to supervise the other network using standard cross-entropy loss, and vice versa. This approach serves two purposes: encouraging high similarity between the predictions of the perturbed networks for the same input image and expanding training data by incorporating unlabeled data with pseudo labels. The experimental results demonstrate that the proposed CPS approach achieves state-of-the-art performance in semi-supervised segmentation on Cityscapes and PASCAL VOC 2012 datasets. Additionally, Figure 3 showcases how our approach outperforms fully-supervised baselines, indicating its effectiveness even when there is a relatively large amount of labeled data available. Furthermore, an ablation study is conducted to evaluate different loss combinations on PASCAL VOC 2012 and Cityscapes datasets. The results consistently show that our approach performs favorably compared to other methods under various partition protocols. Overall, this paper presents a promising solution for improving semi-supervised semantic segmentation by leveraging both labeled and unlabeled data through cross pseudo supervision.