# I2F: A Unified Image-to-Feature Approach for Domain Adaptive Semantic Segmentation

**Authors:** Haoyu Ma, Xiangru Lin, Yizhou Yu  
**arXiv:** [2301.01149](https://arxiv.org/abs/2301.01149)  
**Date:** 2023-01-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.01149v1/

## Key points

- Unsupervised domain adaptation (UDA) in semantic segmentation eliminates the need for extensive annotation efforts.
- Challenges in UDA arise from domain variations in low-level image statistics and high-level contexts, affecting segmentation performance in the target domain.
- The proposed UDA pipeline integrates image-level and feature-level adaptation techniques for semantic segmentation.
- Image-level domain shifts are addressed through global photometric alignment and global texture alignment modules.
- Feature-level domain shifts are handled by performing global manifold alignment of pixel features from both domains onto the source domain's feature manifold.
- Category centers in the source domain are regularized using a category-oriented triplet loss, while target domain consistency regularization is applied over augmented target images.
- Experimental results show a significant improvement over previous methods, with an 8% increase in mean Intersection over Union (mIoU) achieved on the GTA5→Cityscapes task using Deeplab V3+ as the backbone model.
- The proposed method outperforms state-of-the-art techniques by effectively addressing both image-level and feature-level adaptations in UDA for semantic segmentation tasks.

## Summary

, , , , 


In the field of semantic segmentation, unsupervised domain adaptation (UDA) is a crucial task that eliminates the need for extensive annotation efforts. Challenges arise due to domain variations in low-level image statistics and high-level contexts, hindering segmentation performance in the target domain. To address this issue, this paper introduces a novel UDA pipeline for semantic segmentation that integrates both image-level and feature-level adaptation. Specifically, for addressing image-level domain shifts, the proposed approach includes a global photometric alignment module and a global texture alignment module. These modules align images from the source and target domains based on their properties. For handling feature-level domain shifts, the method performs global manifold alignment by projecting pixel features from both domains onto the feature manifold of the source domain. Additionally, category centers in the source domain are regularized using a category-oriented triplet loss, while target domain consistency regularization is applied over augmented target domain images. Experimental results demonstrate significant improvement over previous methods. In a commonly tested GTA5→Cityscapes task utilizing Deeplab V3+ as the backbone model leads to an 8% increase in mean Intersection over Union (mIoU), achieving 58.2%. The proposed method outperforms state-of-the-art techniques by effectively addressing both image-level and feature-level adaptations in UDA for semantic segmentation tasks. Furthermore, qualitative comparisons with existing methods highlight its superiority in various categories such as 'road', 'sidewalk', 'building', 'fence', 'vegetation', 'terrace', 'person', 'car', 'rider', 'truck', 'train', 'bus', 'motor' and 'bike'. is a crucial task that eliminates the need for extensive annotation efforts. Challenges arise due to domain variations in low-level image statistics and high-level contexts, hindering segmentation performance in the target domain. To address this issue, this paper introduces a novel UDA pipeline for  that integrates both  and . Specifically, for addressing image-level domain shifts, the proposed approach includes a global photometric alignment module and a global texture alignment module. These modules align images from the source and target domains based on their properties. For handling feature-level domain shifts, the method performs global manifold alignment by projecting pixel features from both domains onto the feature manifold of the source domain. Additionally, category centers in the source domain are regularized using a category-oriented triplet loss, while target domain consistency regularization is applied over augmented target domain images. Experimental results demonstrate significant improvement over previous methods. In a commonly tested GTA5→Cityscapes task utilizing Deeplab V3+ as the backbone model leads to an 8% increase in mean Intersection over Union (mIoU), achieving 58.2%. The proposed method outperforms state-of-the-art techniques by effectively addressing both image-level and feature-level adaptations in UDA for semantic segmentation tasks. Furthermore, qualitative comparisons with existing methods highlight its superiority in various categories such as 'road', 'sidewalk', 'building', 'fence', 'vegetation', 'terrace', 'person', 'car', 'rider', 'truck', 'train', 'bus', 'motor' and 'bike'.