MLFA: Toward Realistic Test Time Adaptive Object Detection by Multi-Level Feature Alignment.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Object detection models perform well with independent and identically distributed (i.i.d.) data.
• Domain shift between training and testing data significantly degrades object detection performance.
• Test-time adaptive object detection (TTA-OD) offers a realistic solution for online adaptation during testing.

Purpose of the Study:

• To propose a novel multi-level feature alignment (MLFA) method for TTA-OD.
• To enable online adaptation of object detectors using streaming target domain data.
• To improve the robustness and accuracy of object detection in diverse and changing environments.

Main Methods:

• Developed a multi-level feature alignment (MLFA) approach for TTA-OD.
• Selected informative foreground and background features from image feature maps.
• Utilized probabilistic models to capture feature distributions for alignment.
• Implemented global-level feature alignment for domain-invariant features.
• Incorporated cluster-level feature alignment to match category-specific features across domains.

Main Results:

• The proposed MLFA method effectively adapts object detectors to new domains online.
• Global-level alignment encourages the extraction of domain-invariant features.
• Cluster-level alignment aligns category-specific features, improving detection accuracy.
• Extensive experiments demonstrated the effectiveness of the MLFA approach.

Conclusions:

• MLFA enhances TTA-OD by aligning features at multiple levels.
• The method addresses the challenge of domain shift in object detection.
• This approach is crucial for applications like autonomous driving requiring real-time environmental perception.