Domain Adaptive Object Detection via Balancing Between Self-Training and Adversarial Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep learning object detectors face challenges in generalizing to new domains with significant variations.
• Current domain adaptation methods often use image or instance-level adversarial feature alignment, which can be hindered by background noise and lack class-specific alignment.
• Pseudo-labeling for class-level alignment is hampered by noisy predictions due to poor model calibration under domain shift.

Purpose of the Study:

• To develop a technique that balances adversarial feature alignment and class-level alignment for improved object detection domain adaptation.
• To leverage predictive uncertainty to guide the adaptation process, enhancing generalization to new target domains.

Main Methods:

• Quantified predictive uncertainty for both class assignments and bounding-box predictions.
• Utilized low-uncertainty predictions for generating pseudo-labels for self-training.
• Employed high-uncertainty predictions to generate tiles for adversarial feature alignment, focusing on uncertain object regions.

Main Results:

• The proposed method demonstrates a synergistic effect between tiling uncertain regions and pseudo-labeling certain regions, capturing both image and instance-level context.
• Ablation studies confirmed the impact and effectiveness of individual components within the approach.
• Achieved superior performance compared to state-of-the-art methods across five diverse and challenging domain adaptation scenarios.

Conclusions:

• Leveraging predictive uncertainty offers a robust strategy for object detection domain adaptation.
• The proposed method effectively addresses limitations of existing techniques by integrating class-specific and instance-level alignment.
• The approach significantly enhances model generalization capabilities in the face of domain shift.