Strong-Weak Distribution Alignment for Adaptive Object Detection

Abstract: We propose an approach for unsupervised adaptation of object detectors from label-rich to label-poor domains which can significantly reduce annotation costs associated with detection. Recently, approaches that align distributions of source and target images using an adversarial loss have been proven effective for adapting object classifiers. However, for object detection, fully matching the entire distributions of source and target images to each other at the global image level may fail, as domains could have distinct scene layouts and different combinations of objects. On the other hand, strong matching of local features such as texture and color makes sense, as it does not change category level semantics. This motivates us to propose a novel method for detector adaptation based on strong local alignment and weak global alignment. Our key contribution is the weak alignment model, which focuses the adversarial alignment loss on images that are globally similar and puts less emphasis on aligning images that are globally dissimilar. Additionally, we design the strong domain alignment model to only look at local receptive fields of the feature map. We empirically verify the effectiveness of our method on four datasets comprising both large and small domain shifts. Our code is available at \url{https://github.com/VisionLearningGroup/DA_Detection}

• The paper presents a novel method combining weak global and strong local alignment to adapt object detection models effectively across domains.
• It employs selective global matching and local feature consistency, achieving significant mAP improvements on datasets like Pascal VOC to Clipart.
• The approach minimizes annotation costs without sacrificing source domain performance, offering a robust solution for domain adaptation challenges.

Strong-Weak Distribution Alignment for Adaptive Object Detection

The paper presents a method aimed at addressing the challenges associated with unsupervised adaptive object detection, specifically focusing on adapting detectors from label-rich to label-poor domains. This method seeks to mitigate annotation costs, especially when transitioning from well-annotated datasets to those with sparse annotations. The authors propose a novel adaptation strategy that employs strong local alignment along with weak global alignment, thus deviating from conventional approaches that might fail due to varied scene layouts and object combinations inherent in different domains.

Methodology

The core of the proposed approach is two-fold:

1. Weak Global Alignment: This component focuses on partial alignment between domains at the global image level. Unlike methods that enforce full distribution matching, which might be impractical for object detection due to differing scene compositions, this weak alignment selectively emphasizes images that are globally similar. This is achieved by modulating the adversarial alignment loss to prioritize hard-to-classify examples, effectively reducing domain discrepancy without degradation in performance.
2. Strong Local Alignment: Here, the emphasis is on the alignment of local features, such as texture and color, across domains. By utilizing local receptive fields, the method ensures that category-level semantics remain consistent, thereby reducing domain gaps in object appearance without enforcing global image-level invariance.

The approach is empirically tested across four datasets, demonstrating its effectiveness in both significant and minor domain shifts. The authors provide their code openly, enhancing reproducibility and further application by the community.

Numerical Results and Contributions

Across datasets, the proposed method shows a noticeable improvement in mean average precision (mAP) compared to baseline methods. For instance, in the Pascal VOC to Clipart dataset, the proposed model achieved a mAP of 38.1%, significantly outperforming the baseline models that typically score around 25.6%.

The combination of weak global and strong local alignment presents a meaningful approach to minimize domain discrepancy while maintaining critical semantic information necessary for effective object detection. Importantly, this framework does not compromise performance on the source domain, a common issue faced by models enforcing strict domain alignment.

Implications and Future Work

The divergence from full distributional alignment presents a significant contribution to domain adaptation in object detection. This approach implicitly recognizes the heterogeneity inherent in domain shifts, offering a more robust solution compared to traditional models that assume similar domain distributions.

Future developments could explore the integration of this alignment strategy with pixel-level adaptations or other modalities. In particular, advancements in feature extraction and domain classifier architectures could further refine the alignment process. The research hints at potential applications in environments where domain characteristics are subject to frequent change, such as autonomous driving in diverse weather conditions.

In conclusion, the strong-weak distribution alignment methodology offers a pragmatic balance between generalization and specificity, setting a foundation for future exploration in unsupervised domain adaptation for object detection.