Weakly Supervised Adversarial Domain Adaptation for Semantic Segmentation in Urban Scenes.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Semantic segmentation is crucial for pixel-level vision tasks, with Convolutional Neural Networks (CNNs) driving rapid advancements.
• Manual data annotation for training CNNs is labor-intensive and time-consuming.
• Synthetic datasets offer an alternative but exhibit domain discrepancies, hindering performance on real-world data.

Purpose of the Study:

• To develop a weakly supervised adversarial domain adaptation technique for enhancing semantic segmentation from synthetic to real urban scenes.
• To bridge the domain gap between synthetic training data and real-world target domains.

Main Methods:

• A novel three-deep neural network architecture was proposed, comprising a Detection and Segmentation (DS) model, a Pixel-level Domain Classifier (PDC), and an Object-level Domain Classifier (ODC).
• The DS model acts as a generator, while PDC and ODC function as discriminators in an adversarial learning framework.
• The adversarial process aims to train the DS model to learn domain-invariant features, improving generalization.

Main Results:

• The proposed method significantly improves semantic segmentation performance when transferring from synthetic to real urban environments.
• Experimental results demonstrate the effectiveness of the adversarial domain adaptation approach.
• The method established a new record for the mean Intersection over Union (mIoU) metric in this specific problem domain.

Conclusions:

• Weakly supervised adversarial domain adaptation is a viable strategy for overcoming domain shift in semantic segmentation.
• The proposed method effectively learns domain-invariant features, leading to superior performance on real-world data.
• This research contributes to more robust and accurate semantic segmentation models trained on synthetic data.