Adaptive Boosting for Domain Adaptation: Toward Robust Predictions in Scene Segmentation

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptation aims to transfer knowledge from a source to a target domain.
• Current methods often suffer from model bias due to strong source domain supervision.
• Early stopping is used to prevent overfitting, but determining the optimal stopping point is challenging without a target validation set.

Purpose of the Study:

• To propose an efficient bootstrapping method, Adaboost Student, for domain adaptation.
• To eliminate the need for empirical early stopping in domain adaptation.
• To improve the robustness and performance of domain adaptation models.

Main Methods:

• Adaboost Student combines deep model learning with adaptive boosting.
• It employs an adaptive data sampler to focus on hard samples.
• Complementary models are learned during training to prevent overfitting.

Main Results:

• Adaboost Student provides a robust solution without requiring users to determine stopping time.
• The method is orthogonal to existing domain adaptation techniques and can enhance their performance.
• Competitive results were achieved on three scene segmentation domain adaptation benchmarks.

Conclusions:

• Adaboost Student offers an efficient and robust approach to domain adaptation by learning complementary models.
• The method effectively addresses the challenge of early stopping in domain adaptation.
• It shows potential for improving state-of-the-art performance when combined with other domain adaptation strategies.