Abstract
Semi-supervised object detection (SSOD) aims to improve the performance and generalization of existing object detectors by utilizing limited labeled data and extensive unlabeled data. Despite many advances, recent SSOD methods are still challenged by inadequate model refinement using the classical exponential moving average (EMA) strategy, the consensus of Teacher-Student models in the latter stages of training (i.e., losing their distinctiveness), and noisy/misleading pseudo-labels. This paper proposes a novel training-based model refinement (TMR) stage and a simple yet effective representation disagreement (RD) strategy to address the limitations of classical EMA and the consensus problem. The TMR stage of Teacher-Student models optimizes the lightweight scaling operation to refine the model’s weights and prevent overfitting or forgetting learned patterns from unlabeled data. Meanwhile, the RD strategy helps keep these models diverged to encourage the student model to explore additional patterns in unlabeled data. Extensive experiments demonstrate the superior performance of our approach over state-of-the-art SSOD methods. Specifically, the proposed approach outperforms the baseline Unbiased-Teacher-v2 (& Unbiased-Teacher-v1) method by an average mAP margin of 2.23, 2.1, and 3.36 (& 2.07, 1.9 and 3.27) on COCO-standard, COCO-additional, and Pascal VOC datasets, respectively.
Introduction
Object detection models typically require vast amounts of hand-labeled data, a process that is both costly and time-consuming. Semi-supervised object detection (SSOD) addresses this by leveraging a small set of labeled data alongside a large volume of unlabeled images. Most modern SSOD frameworks utilize a "Teacher-Student" architecture where a student model learns from pseudo-labels generated by a teacher. Traditionally, the teacher's weights are updated via an Exponential Moving Average (EMA) of the student's weights.
However, this standard approach suffers from three primary flaws:
- Static Updates: EMA relies on constant smoothing coefficients that do not adapt to the training state.
- The Consensus Problem: Over time, the teacher and student models converge, becoming nearly identical. This lack of distinctiveness prevents the student from learning new patterns.
- Noisy Labels: Errors in the teacher's pseudo-labels can be "memorized" by the student, degrading performance.
Proposed Methodology
The proposed TMR-RD approach introduces two key strategies integrated into a cyclic training pipeline to overcome these limitations.
Training-based Model Refinement (TMR)
Instead of passive EMA updates, the TMR stage introduces a dedicated optimization step. The main network weights are frozen, and lightweight "Scaling Parameters" (SPs) are trained using available labeled data. These SPs adjust the importance of different features, allowing the model to adaptively refine its weights. This process effectively "filters" knowledge transfer, ensuring the teacher does not simply absorb the student's errors.
Representation Disagreement (RD)
To maintain model distinctiveness, the RD strategy calculates the Kullback-Leibler (KL) divergence between the semantic representations of the teacher and student during the semi-supervised learning (SSL) stage. Unlike standard distillation which minimizes this difference, RD ensures the models remain diverged. This forces the student to explore novel patterns in the unlabeled data rather than converging prematurely to the teacher's potentially flawed state.
Experimental Results
The TMR-RD approach was evaluated on industry-standard benchmarks, including MS-COCO and Pascal VOC.
Key findings include:
- State-of-the-Art Performance: On COCO-standard, the method outperformed Unbiased Teacher-v2 by an average of 2.23 mAP. On Pascal VOC, the improvement was even more pronounced at 3.36 mAP.
- Plug-and-Play Versatility: The strategy is architecture-agnostic and was successfully integrated into existing baselines like Unbiased Teacher v1 and v2, yielding immediate performance gains.
- Efficiency and Reliability: The use of lightweight SPs minimizes computational overhead. When combined with cascade regression for bounding box refinement, the model achieved a top score of 50.4 mAP on the COCO-standard dataset using fully-labeled samples.
Conclusion
The TMR-RD approach demonstrates that teacher-update processes should be active optimization problems rather than passive averages. By embracing representation disagreement, the framework avoids the "echo chamber" effect of model consensus. While the TMR stage adds a minor optimization step to the training loop, the performance gains justify the overhead. Future work will explore applying these dynamic refinement techniques to domain adaptation and multi-model knowledge exchange.
Summary updated