# MoDA: Leveraging Motion Priors from Videos for Advancing Unsupervised Domain Adaptation in Semantic Segmentation

**Authors:** Fei Pan, Xu Yin, Seokju Lee, Axi Niu, Sungeui Yoon, In So Kweon  
**arXiv:** [2309.11711](https://arxiv.org/abs/2309.11711)  
**Date:** 2023-09-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.11711v2/

## Key points

- MoDA (Motion-guided Domain Adaptive) framework introduced for unsupervised domain adaptation in semantic segmentation tasks
- Utilizes self-supervised learning techniques to extract object motion cues from unlabeled video frames with geometric constraints
- Aims to facilitate cross-domain alignment by leveraging motion priors for semantic segmentation
- Key components include an object discovery module for localizing and segmenting moving objects, and a semantic mining module for refining pseudo labels based on object masks
- Refined pseudo labels used in self-training loop to bridge domain gap, enhancing annotation quality in target domain
- Experimental results show MoDA outperforms traditional optical flow-based methods in domain alignment effectiveness
- Complements existing state-of-the-art UDA approaches, demonstrating versatility and potential as a valuable addition to the field
- Presented at CVPR 2024 Workshop on Learning with Limited Labelled Data for Image and Video Understanding, received Best Paper Award
- Code implementation available at https://github.com/feipanir/MoDA for further exploration and application

## Summary

In the field of unsupervised domain adaptation (UDA) for semantic segmentation tasks, a novel approach called MoDA (Motion-guided Domain Adaptive) framework has been introduced. This innovative framework addresses the challenge of limited annotations in the target domain by leveraging self-supervised learning techniques to extract object motion cues from unlabeled video frames with geometric constraints. By harnessing these motion priors, MoDA aims to facilitate cross-domain alignment for semantic segmentation. The key components of the MoDA framework include an object discovery module that utilizes object motion information to localize and segment moving objects in the target domain. Subsequently, a semantic mining module refines pseudo labels based on the extracted object masks, enhancing the quality of annotations in the target domain. These refined pseudo labels are then utilized in a self-training loop to bridge the gap between domains. Experimental results on both video and image segmentation tasks demonstrate that MoDA outperforms traditional optical flow-based methods in terms of domain alignment effectiveness. Furthermore, MoDA exhibits versatility by complementing existing state-of-the-art UDA approaches, showcasing its potential as a valuable addition to the field. This research was presented at CVPR 2024 Workshop on Learning with Limited Labelled Data for Image and Video Understanding and received the Best Paper Award. The code implementation of MoDA is available at https://github.com/feipanir/MoDA, providing a valuable resource for further exploration and application of this innovative framework.