Abstract
Cross-domain object detection remains a significant challenge due to domain shift between training and deployment environments. While unsupervised domain adaptation (UDA) via adversarial learning is a common solution, the inherent instability of the adversarial "game" often leads to sub-optimal performance. This paper introduces BDC-FR, a framework that stabilizes adversarial training through a Balanced Domain Classifier (BDC) and a dynamic learning rate strategy. By optimizing the classifier's architecture to match the feature extractor's capacity and enforcing multi-level consistency, BDC-FR achieves state-of-the-art results across various benchmarks, including adverse weather and synthetic-to-real transitions.
Introduction
Object detection models, such as Faster R-CNN, typically suffer from performance degradation when deployed in environments that differ from their training data—a phenomenon known as domain shift. For instance, a model trained on clear-weather urban scenes may fail in foggy or rainy conditions. Unsupervised Domain Adaptation (UDA) aims to bridge this gap by aligning feature distributions between a labeled source domain and an unlabeled target domain.
Adversarial learning is a prominent UDA technique where a domain classifier and a feature extractor compete. However, maintaining equilibrium in this competition is difficult; a classifier that is too strong or too weak can hinder the feature extractor's ability to learn domain-invariant representations. BDC-FR addresses this by ensuring the domain classifier and feature extractor remain in equilibrium through architectural balancing and dynamic training strategies.
Methods
The BDC-FR framework introduces three core innovations to stabilize adversarial training and improve domain alignment.
Balanced Domain Classifier (BDC)
The primary innovation is the BDC, which moves beyond generic classifier architectures. The authors propose an Iterative Control Variable Method to systematically determine the optimal architecture for the domain classifier. By tuning hyperparameters such as the number of layers and kernel sizes, the BDC's learning capacity is calibrated to match the feature extractor. This balance prevents the vanishing gradient problem and ensures that the adversarial process remains productive throughout training.
Double-Consistency Regularization (DCR)
To ensure robust alignment, BDC-FR operates at three distinct levels: image-level, instance-level (individual objects), and region-level (proposals from the Region Proposal Network). DCR enforces consistency between these levels. For example, if the image-level classifier identifies a sample as belonging to the target domain, the instance-level and region-level classifiers should ideally yield the same prediction. This hierarchical alignment ensures that domain invariance is captured at both global and local scales.
Dynamic Learning Rate Strategy
To further stabilize training, a piecewise dynamic learning rate strategy is employed. A higher learning rate is used initially to accelerate convergence when prediction errors are large. As the model approaches an equilibrium point, the learning rate is reduced to allow for fine-grained optimization and to prevent the model from oscillating around the optimal solution.
Experimental Evaluation
The effectiveness of BDC-FR was evaluated on standard cross-domain benchmarks: Cityscapes to Foggy Cityscapes (adverse weather), SIM 10k to Cityscapes (synthetic-to-real), and Cityscapes to KITTI (cross-camera).
Adverse Weather Adaptation
On the Foggy Cityscapes dataset, BDC-FR achieved a mean Average Precision (mAP) of 42.9%, outperforming the previous state-of-the-art (TIA) and significantly improving upon the base Faster R-CNN model (22.6% mAP). The results across various object categories demonstrate the robustness of the balanced approach.
Synthetic-to-Real and Cross-Camera Scenarios
In the SIM 10k to Cityscapes task, BDC-FR reached 45.3% AP for the "car" category, surpassing MeGA-CDA (44.8%). For the cross-camera adaptation (Cityscapes to KITTI), the model achieved 44.1% mAP, slightly exceeding the performance of DIR-FR (44.0%). Ablation studies further confirmed that removing the BDC component leads to a significant drop in performance (from 42.9% to 39.2% mAP), highlighting its critical role.
Conclusion
BDC-FR demonstrates that the structural capacity of the domain classifier is a critical factor in successful unsupervised domain adaptation. By balancing the adversarial game and enforcing multi-level consistency, the framework provides a stable and efficient path to domain-invariant object detection. Future research may explore applying these principles to Transformer-based architectures and more complex open-set domain adaptation tasks.
Summary updated