Abstract
Semi-supervised object detection (SSOD) leverages small amounts of labeled data alongside large unlabeled datasets. The prevailing Teacher-Student framework often plateaus due to static Exponential Moving Average (EMA) updates and premature model convergence. We present TMR-RD, which introduces a Training-based Model Refinement (TMR) stage to dynamically evolve the teacher model and a Representation Disagreement (RD) strategy to preserve diversity between the teacher and student. Our method achieves state-of-the-art performance on MS-COCO and Pascal VOC benchmarks, significantly outperforming existing baselines.
Introduction
Semi-supervised object detection (SSOD) aims to train robust detectors by utilizing a limited set of labeled images and a vast pool of unlabeled data. The standard approach involves a Teacher-Student framework: a "Teacher" model generates pseudo-labels for unlabeled images, which then serve as ground truth to train a "Student" model. Despite its success, this framework faces two critical bottlenecks:
- The EMA Trap: Traditionally, the Teacher is updated via an Exponential Moving Average (EMA) of the Student’s weights. This relies on a fixed, manually-tuned smoothing coefficient that fails to adapt to the training progress, often resulting in noisy pseudo-labels.
- The Consensus Problem: Over time, the Student and Teacher models tend to converge and become identical. When they reach perfect agreement, the Student stops learning new information from the Teacher, causing performance to plateau prematurely.
The TMR-RD Approach
To overcome these limitations, we propose two primary innovations: Training-based Model Refinement (TMR) and Representation Disagreement (RD).
Training-based Model Refinement (TMR)
Instead of a static EMA update, we introduce a dedicated TMR stage. Inspired by meta-transfer learning, we freeze the primary model weights and optimize a set of lightweight Scaling Parameters (SPs). These parameters learn to aggregate information from both labeled and unlabeled data effectively. After optimizing these SPs, the model weights are refined using a dynamic update rule, allowing the Teacher to evolve more intelligently than a simple moving average allows.
Representation Disagreement (RD)
To prevent premature convergence, the RD strategy encourages the models to maintain a healthy level of diversity. During the semi-supervised learning (SSL) stage, we calculate the Kullback-Leibler (KL) divergence between the high-level semantic representations of the Teacher and Student. By incorporating this as a loss term, we force the Student to explore more diverse patterns in the unlabeled data rather than simply mimicking the Teacher.
The overall training follows a cyclic pattern: Burn-In (initial supervised training) $\rightarrow$ SSL (Student learning with RD) $\rightarrow$ TMR (Teacher refinement) $\rightarrow$ Repeat.
Experimental Results
We evaluated TMR-RD on the MS-COCO and Pascal VOC benchmarks. The results demonstrate consistent improvements over state-of-the-art (SOTA) methods across various data regimes.
Key Findings
- MS-COCO Standard: On the 10% labeled split, TMR-RD-v2 achieved 40.93 mAP, a significant improvement over the Unbiased Teacher-v2 baseline.
- MS-COCO Additional: Utilizing the full labeled set plus 123k unlabeled images, the model reached 50.4 mAP, outperforming the baseline by 5.6 points.
- Pascal VOC: The method showed a massive improvement, surpassing the baseline by an average mAP margin of 3.36.
- Ablation Studies: Our experiments confirmed that both TMR and RD contribute independently to the performance gains; removing either component leads to a noticeable drop in accuracy.
Conclusion
TMR-RD represents a shift from rigid, manually-tuned update rules toward dynamic, optimization-based model refinement in SSOD. By treating the teacher update as a learning problem and explicitly maintaining model diversity through representation disagreement, we break the performance plateaus common in traditional Teacher-Student frameworks. This training strategy is "plug-and-play," showing successful integration with various underlying detectors like Faster R-CNN and Cascade R-CNN.
Summary updated