Source-linked AI summary
Unbiased Teacher v2: Semi-supervised Object Detection for Anchor-free and Anchor-based Detectors
Yen-Cheng Liu, Chih-Yao Ma, Zsolt Kira
TL;DR
Semi-supervised object detection still lacked evidence for anchor-free detectors and effective bounding-box regression pseudo-labeling. Unbiased Teacher v2 studies these gaps and introduces Listen2Student, which selects regression pseudo-labels using relative Teacher–Student uncertainties. Across anchor-free and anchor-based detectors on VOC and COCO benchmarks, the method consistently performs favorably against state-of-the-art approaches.
Problem
Prior semi-supervised object detection work had not addressed anchor-free detectors effectively and was ineffective for pseudo-labeling bounding-box regression.
Method
Unbiased Teacher v2 generalizes SS-OD to anchor-free detectors and uses Listen2Student to select boundary-level regression pseudo-labels when Teacher uncertainty is lower than Student uncertainty.
Results
Across VOC, COCO-standard, and COCO-additional, the method consistently performs favorably against state-of-the-art methods for both anchor-free and anchor-based detectors.
Takeaways & Limitations
Relative Teacher–Student uncertainty can prevent misleading regression pseudo-labels and support improvements in the regression branch under semi-supervised training.
Takeaways & Limitations
Teacher confidence alone cannot explicitly prevent regression pseudo-labels whose predicted direction contradicts the ground-truth direction.
Abstract
from arXiv · showhide
With the recent development of Semi-Supervised Object Detection (SS-OD) techniques, object detectors can be improved by using a limited amount of labeled data and abundant unlabeled data. However, there are still two challenges that are not addressed: (1) there is no prior SS-OD work on anchor-free detectors, and (2) prior works are ineffective when pseudo-labeling bounding box regression. In this paper, we present Unbiased Teacher v2, which shows the generalization of SS-OD method to anchor-free detectors and also introduces Listen2Student mechanism for the unsupervised regression loss. Specifically, we first present a study examining the effectiveness of existing SS-OD methods on anchor-free detectors and find that they achieve much lower performance improvements under the semi-supervised setting. We also observe that box selection with centerness and the localization-based labeling used in anchor-free detectors cannot work well under the semi-supervised setting. On the other hand, our Listen2Student mechanism explicitly prevents misleading pseudo-labels in the training of bounding box regression; we specifically develop a novel pseudo-labeling selection mechanism based on the Teacher and Student's relative uncertainties. This idea contributes to favorable improvement in the regression branch in the semi-supervised setting. Our method, which works for both anchor-free and anchor-based methods, consistently performs favorably against the state-of-the-art methods in VOC, COCO-standard, and COCO-additional.
1. Introduction
Unbiased Teacher v2 addresses two gaps in semi-supervised object detection: limited validation on anchor-free detectors and ineffective regression pseudo-labeling. It generalizes SS-OD across detector types and introduces Listen2Student to select regression pseudo-labels using relative Teacher–Student uncertainty.
- Motivation: Existing SS-OD methods improve anchor-free detectors less than anchor-based models because centerness-based selection and localization-based labeling are unreliable with limited supervision.The study specifically identifies unreliable centerness scores and sensitivity to localization noise in pseudo-labels.
- Motivation: Prior regression pseudo-labeling methods inadequately represent the quality of four predicted boundaries and may use classification confidence that does not reflect regression quality.The proposed approach instead predicts regression uncertainty and uses relative Teacher–Student uncertainty for boundary-level selection.
- Method: Listen2Student compares Teacher and Student regression uncertainties and applies unsupervised regression loss only when the Teacher has lower uncertainty.This selection targets misleading regression pseudo-labels rather than relying on a single confidence measure.
- Evaluation: The method is evaluated on COCO-standard, COCO-additional, and VOC benchmarks, with ablations examining Listen2Student.The paper reports significant improvements against state-of-the-art SS-OD methods on both detector types.
- Contributions: The method generalizes semi-supervised object detection to both anchor-free and anchor-based detectors, identifying core pseudo-labeling issues in anchor-free models.The authors report favorable improvements against state-of-the-art methods and describe analyses across both detector families.
2. Related Work
The related work distinguishes anchor-based and anchor-free object detectors and reviews semi-supervised detection methods based on augmentation, consistency regularization, and pseudo-labeling. It motivates testing whether approaches developed mainly for anchor-based detectors transfer effectively to anchor-free models.
- Anchor-Free Object Detectors: Anchor-based detectors predict shifts and scales relative to predefined anchor boxes, whereas anchor-free detectors remove predefined anchors and predict box geometry directly.The paper describes anchor-free representations using points or per-pixel boundary distances and centerness.
- Anchor-Free Object Detectors: FCOS predicts classification scores, distances to four boundaries, and centerness for each pixel, and serves as the paper’s base anchor-free model.The paper positions FCOS among methods improving anchor-free detection through adaptive selection, soft labels, pyramid-level selection, and boundary uncertainty modeling.
- Semi-Supervised Object Detection: Semi-supervised learning methods use augmentations or perturbations with consistency regularization to exploit unlabeled images alongside limited labeled data.The related work connects this framework to semi-supervised object detection methods using pseudo-labels.
3. Method
Unbiased Teacher v2 adapts semi-supervised object detection to anchor-free detectors by diagnosing pseudo-labeling failures and introducing Listen2Student for regression supervision.
- Pseudo-labeling on Anchor-Free Detectors: Existing pseudo-labeling methods provide much smaller improvements on anchor-free detectors than on anchor-based detectors in semi-supervised learning.The study attributes this gap to centerness bias and unreliable label assignment under localization noise.
- Pseudo-labeling on Anchor-Free Detectors: Selecting pseudo-boxes by classification score outperforms box-score selection in semi-supervised anchor-free detection because centerness can dominate unreliable box scores.Anchor-free box scores multiply classification and centerness scores, allowing high-centerness, low-classification predictions to be selected.
- Pseudo-labeling on Anchor-Free Detectors: Standard label assignment is more robust than center sampling and localization-based labels when pseudo-boxes contain localization noise.The cited evaluation reports lower pixel-wise recall and precision for the specialized alternatives under noisy pseudo-boxes.
- Listen2Student: Listen2Student selects regression supervision at the boundary level using relative Teacher–Student uncertainty, retaining boundaries where the Teacher is less uncertain.The mechanism removes misleading regression instances rather than relying only on Teacher confidence.
- Listen2Student: For unlabeled data, regression uncertainty is predicted with an additional branch because ground-truth regression labels are unavailable.This uncertainty is intended to loosely correlate with regression error and supports boundary-level pseudo-label selection.
4. Experiments
Experiments evaluate Unbiased Teacher v2 across anchor-free and anchor-based detectors on COCO and VOC settings, including regression-loss ablations. The method consistently improves over baselines and benefits precise boundary prediction.
- Experimental Setup: The experiments use COCO-standard labeled fractions from 0.5% to 10%, plus COCO-additional and VOC scenarios, with FCOS and Faster R-CNN evaluations.Implementation uses 8 labeled and 8 unlabeled images per batch unless otherwise specified, with λu = 3.0 and classification threshold τ = 0.5.
- Results across benchmarks: Across COCO-standard, VOC, and COCO-additional, the method consistently outperforms or favorably compares with existing semi-supervised detection baselines.The experiments cover both anchor-free and anchor-based detectors, with COCO-standard, COCO-additional, and VOC scenarios.
- Results on Anchor-free Detector: On the anchor-free FCOS setting, the method performs favorably against adapted CSD, STAC, and Unbiased Teacher baselines across supervision levels.Each method is run five times, with means and variances reported for COCO-standard.
- Results across benchmarks: 35.08 mAP is achieved on COCO-standard with 10% labeled data using the anchor-based detector, while reducing the anchor-free versus anchor-based performance gap.The improvement remains under different batch sizes, and the reported result uses the proposed framework on Faster R-CNN.
- Results across benchmarks: 58.08 mAP is achieved on VOC when VOC12 and COCO20cls are used as unlabeled data, while COCO2017-unlabeled yields 44.75 mAP.Using VOC12 alone as unlabeled data yields 56.87 mAP; the COCO-additional result is reported without inference-threshold tuning.
- Effectiveness of Unsupervised Regression Loss: Confidence-thresholded regression improves easier metrics such as AP55 but can degrade AP95, whereas Listen2Student improves consistently across evaluation metrics.The ablation changes only the unsupervised regression loss; Listen2Student uses boundary-wise uncertainty comparisons for pseudo-label selection.
- Limitations: Large-scale unlabeled datasets, localization-uncertainty estimation, unseen objects, and domain shift remain open challenges for the approach.The paper identifies these as limitations and future-work directions rather than evaluated components of the experiments.
5. Conclusion
The conclusion presents Unbiased Teacher v2 as a semi-supervised detection framework evaluated on anchor-free and anchor-based models. Its Listen2Student mechanism uses relative Teacher–Student uncertainty to improve regression pseudo-labeling, while large-scale data and domain shift remain challenges.
- 5. Conclusion: The study establishes anchor-free semi-supervised detection benchmarks and examines existing methods alongside the proposed framework.This extends the evaluation scope beyond the anchor-based detectors emphasized by prior work.
- 5. Conclusion: Listen2Student selects beneficial boundary-wise regression pseudo-labels by comparing relative Teacher and Student uncertainties.The mechanism explicitly prevents misleading regression pseudo-labels and enables the regression branch to use unlabeled images.
- 5. Conclusion: Across three semi-supervised detection tasks, experiments report consistent improvements and verify generalization to both anchor-free and anchor-based detectors.The conclusion summarizes an extensive study of Listen2Student's effectiveness and generalization.