Robust Domain Adaptive Object Detection With Unified Multi-Granularity Alignment.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Domain adaptive detection seeks to enhance object detector generalization to new target domains.
• Current methods align features across domains using adversarial learning but often neglect inter-granularity relationships.
• This oversight can degrade detection performance by failing to capture complex feature dependencies.

Purpose of the Study:

• To propose a unified multi-granularity alignment (MGA) framework for domain-invariant feature learning in object detection.
• To address limitations in existing methods by simultaneously encoding dependencies across pixel, instance, and category levels.
• To improve the robustness and generalization capabilities of detectors in domain adaptive scenarios.

Main Methods:

• Developed an omni-scale gated fusion (OSGF) module for aggregating instance representations from pixel-level features using scale-aware convolutions.
• Introduced multi-granularity discriminators to identify the domain origin (source or target) of features at different levels.
• Implemented an adaptive exponential moving average (AEMA) strategy for model updates, improving pseudo-labels and mitigating local misalignment.

Main Results:

• The MGA framework effectively learns domain-invariant features by simultaneously aligning multiple granularities.
• Experiments demonstrated the superiority of MGA on FCOS and Faster R-CNN detectors across various domain adaptation scenarios.
• The proposed OSGF module and AEMA strategy contributed to robust multi-scale detection and alleviated misalignment issues.

Conclusions:

• The unified multi-granularity alignment (MGA) framework offers a novel approach to domain adaptive detection.
• Simultaneous alignment across pixel, instance, and category levels significantly enhances detector generalization.
• MGA provides a robust and effective solution for improving object detection performance in domain adaptive settings.