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

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) for semantic segmentation aims to reduce manual annotation efforts.
• Domain discrepancies in low-level statistics and high-level contexts hinder target domain performance.
• Existing UDA methods often lack unified approaches for both image-level and feature-level adaptation.

Purpose of the Study:

• To propose a novel UDA pipeline that unifies image-level and feature-level adaptation for semantic segmentation.
• To address domain shifts by aligning images and features across source and target domains.
• To improve segmentation performance in the target domain without manual annotations.

Main Methods:

• Developed a unified UDA pipeline integrating global photometric and texture alignment modules for image-level adaptation.
• Implemented global manifold alignment by projecting pixel features onto the source domain's feature manifold for feature-level adaptation.
• Utilized a category-oriented triplet loss for source domain category center regularization and target domain consistency regularization.

Main Results:

• The proposed pipeline significantly outperforms previous UDA methods for semantic segmentation.
• Achieved a state-of-the-art (SOTA) performance improvement of 8% on the GTA5 → Cityscapes task.
• Reached 58.2% mean Intersection over Union (mIoU) using Deeplab V3+ as the backbone.

Conclusions:

• The unified approach effectively tackles domain discrepancies in semantic segmentation.
• Joint image-level and feature-level adaptation is crucial for robust UDA.
• The proposed method sets a new benchmark for unsupervised domain adaptation in semantic segmentation.