AFAN: Augmented Feature Alignment Network for Cross-Domain Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation for object detection is crucial for real-world applications but underexplored.
• Existing methods struggle with limited annotated data and insufficient feature alignment for domain-invariant representations.

Purpose of the Study:

• To propose a novel augmented feature alignment network (AFAN) to address limitations in unsupervised domain adaptation for object detection.
• To enhance feature alignment and learn domain-invariant representations by integrating intermediate domain image generation and domain-adversarial training.

Main Methods:

• Developed an augmented feature alignment network (AFAN) integrating intermediate domain image generation and domain-adversarial training.
• Introduced an intermediate domain image generator with soft domain labels to bridge domain divergence and augment data.
• Implemented feature pyramid alignment and region feature alignment with discriminators to align multi-scale and proposal features.

Main Results:

• The proposed AFAN significantly outperforms state-of-the-art methods on standard benchmarks for both similar and dissimilar domain adaptations.
• Extensive experiments validated the effectiveness of individual components within the AFAN framework.
• Demonstrated the network's capability to learn robust domain-invariant representations for object detection.

Conclusions:

• The novel AFAN effectively tackles challenges in unsupervised domain adaptation for object detection.
• The integration of intermediate domain image generation and advanced feature alignment strategies leads to superior performance.
• AFAN provides a promising direction for developing more accurate and generalizable object detection models across different domains.