SSDA-YOLO: Semi-supervised Domain Adaptive YOLO for Cross-Domain Object Detection (2211.02213v2)

Abstract: Domain adaptive object detection (DAOD) aims to alleviate transfer performance degradation caused by the cross-domain discrepancy. However, most existing DAOD methods are dominated by outdated and computationally intensive two-stage Faster R-CNN, which is not the first choice for industrial applications. In this paper, we propose a novel semi-supervised domain adaptive YOLO (SSDA-YOLO) based method to improve cross-domain detection performance by integrating the compact one-stage stronger detector YOLOv5 with domain adaptation. Specifically, we adapt the knowledge distillation framework with the Mean Teacher model to assist the student model in obtaining instance-level features of the unlabeled target domain. We also utilize the scene style transfer to cross-generate pseudo images in different domains for remedying image-level differences. In addition, an intuitive consistency loss is proposed to further align cross-domain predictions. We evaluate SSDA-YOLO on public benchmarks including PascalVOC, Clipart1k, Cityscapes, and Foggy Cityscapes. Moreover, to verify its generalization, we conduct experiments on yawning detection datasets collected from various real classrooms. The results show considerable improvements of our method in these DAOD tasks, which reveals both the effectiveness of proposed adaptive modules and the urgency of applying more advanced detectors in DAOD. Our code is available on \url{https://github.com/hnuzhy/SSDA-YOLO}.

• The paper introduces a novel SSDA-YOLO framework that integrates YOLOv5 with semi-supervised domain adaptation to tackle domain shift challenges.
• It employs a Mean Teacher model for knowledge distillation and utilizes scene style transfer to align image-level features between source and target domains.
• Experimental results on benchmarks like Cityscapes and Foggy Cityscapes demonstrate significant mAP improvements over traditional two-stage detectors.

An Analysis of SSDA-YOLO for Cross-Domain Object Detection

The paper entitled "SSDA-YOLO: Semi-supervised Domain Adaptive YOLO for Cross-Domain Object Detection" introduces an innovative methodology for improving cross-domain object detection, primarily addressing limitations associated with the commonly used two-stage detector, Faster R-CNN, in the domain adaptive object detection (DAOD) field. Instead, the authors leverage the single-stage YOLOv5 detector, known for its efficiency and near real-time performance, making the method potentially more viable for practical applications.

The motivation behind this paper stems from the need to address the performance degradation when object detection models, trained on a source domain, are applied to a significantly different target domain. This domain shift is often characterized by variations such as dissimilar imaging styles, lighting conditions, and perspectives. Given that the Faster R-CNN detector, despite its accuracy, is computationally intensive and dominated previous DAOD studies, the proposal to use YOLOv5 is timely and highly appropriate for resource-sensitive environments.

Methodology Summary

The authors propose a semi-supervised domain adaptive YOLO (SSDA-YOLO) detector, which is constructed around a few core innovations:

1. Knowledge Distillation with Mean Teacher Model: The SSDA-YOLO framework uses a knowledge distillation approach, employing a teacher-student setup via the Mean Teacher model. This configuration enables efficient extraction of instance-level features from the unlabeled target domain, enhancing the detector's ability to learn domain-invariant representations.
2. Scene Style Transfer: To mitigate discrepancies at the image level, the paper introduces the use of pseudo images generated through scene style transfer. This process helps align image-level features across the source and target domains and ensures that the style variance is addressed efficiently.
3. Consistency Loss: A novel consistency loss function is formulated to harmonize predictions across domains, further ensuring robust cross-domain prediction alignment.

Experimental Evaluation

The authors validate their approach on multiple public benchmarks, including PascalVOC, Clipart1k, Cityscapes, and Foggy Cityscapes, as well as real-world application scenarios such as yawning detection in classroom settings. The SSDA-YOLO framework shows significant performance improvements over baseline detectors, and comparisons with state-of-the-art methods denote competitive results.

Numerical Results:

• The SSDA-YOLO method achieves a substantial gain in mAP over existing systems like SWDA and TIA, demonstrating both robustness and viability.
• For example, the method records notable mAP enhancements, particularly in the transfer case of Cityscapes to Foggy Cityscapes, reflecting the effectiveness of the proposed adaptive mechanisms across adverse weather conditions.

Practical and Theoretical Implications

Practically, the deployment of SSDA-YOLO with YOLOv5 addresses the urgency of adopting more advanced, faster detectors in the DAOD landscape, bridging the gap between industrial applicability and academic rigor. Theoretically, the paper challenges the established reliance on outdated methodologies in DAOD, advocating for a shift towards single-stage detectors complemented by strategic domain adaptation techniques.

Future Directions

This paper opens avenues for future research into enhancing DAOD frameworks with advanced, lightweight detectors like YOLOv5, encouraging further exploration into semi-supervised learning paradigms. The incorporation of transformer-based architectures and self-supervised learning could potentially elevate the robustness of domain adaptive detection models.

In conclusion, the proposed SSDA-YOLO offers a compelling case for leveraging contemporary detectors in DAOD tasks, presenting a balanced approach to achieving high performance in cross-domain object detection.