# A Re-Parameterized Vision Transformer (ReVT) for Domain-Generalized Semantic Segmentation

**Authors:** Jan-Aike Termöhlen, Timo Bartels, Tim Fingscheidt  
**arXiv:** [2308.13331](https://arxiv.org/abs/2308.13331)  
**Date:** 2023-08-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.13331v1/

## Key points

- Semantic segmentation involves assigning semantic labels to each pixel of an image
- Existing models tend to perform poorly in unseen domains with different data distributions
- Authors propose an augmentation-driven approach for domain generalization in semantic segmentation
- Method utilizes a re-parameterized vision transformer (ReVT) and weight averaging of multiple models
- Approach achieves state-of-the-art mean Intersection over Union (mIoU) performance on benchmark datasets
- Reported mIoU scores are 47.3% for small models and 50.1% for midsized models, surpassing prior art results
- Proposed method requires fewer parameters and achieves higher frame rate compared to previous techniques
- It is easy to implement and does not introduce computational complexity during inference like network ensembles

## Summary

The task of semantic segmentation involves assigning semantic labels to each pixel of an image. However, the performance of existing models tends to decline when deployed in an unseen domain with different data distributions than the training domain. To address this issue, the authors propose a novel augmentation-driven approach for domain generalization in semantic segmentation. Their method utilizes a re-parameterized vision transformer (ReVT) and incorporates weight averaging of multiple models after training. The authors evaluate their approach on various benchmark datasets and achieve state-of-the-art mean Intersection over Union (mIoU) performance. Specifically, they report mIoU scores of 47.3% for small models and 50.1% for midsized models, surpassing prior art results of 46.3% and 47.8%, respectively, on commonly used benchmark datasets. Notably, the proposed method offers several advantages over previous techniques. It requires fewer parameters and achieves a higher frame rate compared to the best prior art solution. Additionally, it is easy to implement and does not introduce any computational complexity during inference, unlike network ensembles. In summary, the authors present a re-parameterized vision transformer-based approach for domain-generalized semantic segmentation that outperforms previous methods in terms of mIoU performance while maintaining efficiency and simplicity in implementation.