A Curriculum Domain Adaptation Approach to the Semantic Segmentation of Urban Scenes

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel at semantic segmentation but require extensive annotated data.
• Training CNNs on synthetic data is hindered by the domain mismatch between synthetic and real-world images.

Purpose of the Study:

• To propose a curriculum-style learning approach to minimize the domain gap in urban scene semantic segmentation.
• To improve the performance of CNNs trained on synthetic data for real-world applications.

Main Methods:

• A curriculum domain adaptation strategy is employed, starting with easy tasks like estimating global and local label distributions.
• The method infers properties of the target domain (urban scenes) from their inherent characteristics.
• A segmentation network is trained while regularizing predictions to align with inferred properties.

Main Results:

• The proposed method demonstrates superior performance compared to baseline approaches.
• Experiments were conducted on two datasets using three different backbone networks.
• Ablation studies confirmed the effectiveness of the approach.

Conclusions:

• Curriculum domain adaptation is effective in reducing the domain gap for urban scene semantic segmentation.
• The method offers a viable solution for training robust segmentation models with limited real-world annotated data.