Dual Decoupling Training for Semi-supervised Object Detection with Noise-Bypass Head

Shida Zheng, Chenshu Chen, Xiaowei Cai, Tingqun Ye, Wenming TanAAAI 2022 · DOI

Introduction

Semi-supervised object detection (SSOD) leverages a small amount of labeled data alongside a large pool of unlabeled data to improve detector performance. A common approach is the "teacher-student" paradigm, where a teacher model generates pseudo-labels on unlabeled images to supervise a student model. However, this method faces a fundamental "noise vs. recall" trade-off. High confidence thresholds ensure "clean" labels but result in low recall, missing difficult or small objects. Conversely, low thresholds increase recall but introduce significant noise (false positives), which can degrade the student model's precision.

Furthermore, traditional SSOD methods often rely solely on classification confidence to filter pseudo-labels. This overlooks the potential misalignment between classification and localization: a model may be certain about an object's category but imprecise about its bounding box coordinates.

The Dual Decoupling Training (DDT) Framework

To address these challenges, the authors propose Dual Decoupling Training (DDT). This framework introduces two levels of decoupling to effectively utilize noisy pseudo-labels without compromising the main detector's accuracy.

Conceptual diagram of the dual decoupling strategy, illustrating how samples are categorized by classification and localization scores.

Data Decoupling: The Noise-Bypass Head

DDT employs a two-level thresholding strategy to categorize pseudo-boxes into three groups: clean foreground, noisy foreground, and clean background. Instead of discarding noisy foreground samples, DDT introduces a secondary, parallel ROI head called the Noise-Bypass Head.

  • Main Head: Trained exclusively on "clean" foreground samples to maintain high precision.
  • Noise-Bypass Head: Designed to "absorb" noisy foreground samples. By training on these diverse but uncertain labels, the shared backbone network learns richer feature representations.

This architecture is inspired by bypass capacitors in electronics, which filter out high-frequency noise to stabilize voltage. Here, the bypass head prevents noisy labels from polluting the main head's training signal while still allowing the model to benefit from the information they contain.

Task Decoupling: Classification vs. Localization

DDT treats classification and localization as independent qualities. It calculates a localization score based on the consistency (IoU) between the predictions of the Main Head and the Noise-Bypass Head. If both heads agree on a location, the box is deemed well-localized, regardless of its classification confidence. This decoupling ensures that high-quality localization information is not discarded simply because of lower classification certainty.

The DDT architecture featuring the Teacher-Student EMA framework with parallel Main and Noise-Bypass heads.

Training and Augmentation

The framework utilizes an Exponential Moving Average (EMA) update scheme for the teacher-student models. To further enhance robustness and prevent the dual-head structure from overfitting, heavy "inter-data" augmentations like Mixup and Mosaic are employed. These techniques fuse multiple images, providing a more complex and varied training context.

Key Results

DDT achieves state-of-the-art performance on major benchmarks:

  • PASCAL VOC: Reached 54.7 mAP, a 1.7-point improvement over previous methods like Humble Teacher.
  • MS-COCO: Consistently outperformed existing SOTA methods by approximately 1.0 mAP across various labeling protocols (from 1% to 10% labeled data).
  • Localization Precision: Qualitative analysis confirms that DDT produces tighter, more accurate bounding boxes compared to baseline models, validating the effectiveness of the independent localization scoring.

Conclusion

Dual Decoupling Training redefines noisy data as a valuable resource rather than a nuisance. By decoupling clean data from noisy data via a specialized bypass head, and classification from localization, DDT extracts diverse information from unlabeled data while maintaining high precision. Importantly, the Noise-Bypass head is discarded during inference, ensuring that the final model remains as efficient as a standard detector.

Summary updated

BibTeX
@inproceedings{zheng2022dual,
  author = {Shida Zheng and Chenshu Chen and Xiaowei Cai and Tingqun Ye and Wenming Tan},
  title = {{Dual Decoupling Training for Semi-supervised Object Detection with Noise-Bypass Head}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i3.20264},
}
Figure