PaperlayerView exact revision on arXiv

Source-linked AI summary

Unbiased Teacher for Semi-Supervised Object Detection

Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, Peter Vajda

arXiv:2102.09480v1cs.CVcs.LG

TL;DR

Semi-supervised object detection remains less explored than image classification despite the cost of bounding-box annotation, and pseudo-labeling can be biased by class imbalance. Unbiased Teacher jointly trains a student and an EMA-updated teacher, using class-balancing mechanisms to improve pseudo-labels; it delivers substantial gains across COCO and VOC benchmarks. The approach does not apply unsupervised losses to bounding-box regression because confidence thresholding does not measure localization quality.

  • Problem

    Semi-supervised object detection is underexplored relative to image classification, while scarce labels and class imbalance create overfitting and pseudo-labeling bias.

  • Method

    Unbiased Teacher jointly trains a Student from Teacher-generated pseudo-labels and gradually updates the Teacher from the Student via EMA, using different augmentations and Focal loss.

  • Results

    Unbiased Teacher achieves state-of-the-art performance across COCO-standard, COCO-additional, and VOC, including a 6.8 absolute mAP improvement over STAC with 1% labeled MS-COCO data.

  • Takeaways & Limitations

    The framework provides a unified Teacher-Student approach that improves pseudo-label quality while addressing overfitting and class-imbalance issues in low-label object detection.

  • Takeaways & Limitations

    The method omits unsupervised bounding-box regression losses because predicted-box confidence reflects category confidence rather than localization quality.

Abstract

from arXiv · show

Semi-supervised learning, i.e., training networks with both labeled and unlabeled data, has made significant progress recently. However, existing works have primarily focused on image classification tasks and neglected object detection which requires more annotation effort. In this work, we revisit the Semi-Supervised Object Detection (SS-OD) and identify the pseudo-labeling bias issue in SS-OD. To address this, we introduce Unbiased Teacher, a simple yet effective approach that jointly trains a student and a gradually progressing teacher in a mutually-beneficial manner. Together with a class-balance loss to downweight overly confident pseudo-labels, Unbiased Teacher consistently improved state-of-the-art methods by significant margins on COCO-standard, COCO-additional, and VOC datasets. Specifically, Unbiased Teacher achieves 6.8 absolute mAP improvements against state-of-the-art method when using 1% of labeled data on MS-COCO, achieves around 10 mAP improvements against the supervised baseline when using only 0.5, 1, 2% of labeled data on MS-COCO.

1 INTRODUCTION

Semi-supervised object detection addresses costly bounding-box annotation by training with limited labeled data and abundant unlabeled data. Unbiased Teacher targets pseudo-labeling bias and overfitting, achieving strong results across several benchmarks.

  • Object detection has received less semi-supervised attention than image classification because bounding-box annotations require more effort.
  • Unbiased Teacher jointly trains a Student and a slowly progressing Teacher, with the Teacher generating pseudo-labels and the Student updating the Teacher via EMA.The models receive different augmented input images.
  • The framework uses pseudo-labels as supervision for RPN and ROIhead and combines EMA with Focal loss to address overfitting and class-imbalance bias.
  • 6.8 absolute mAP improvement over STAC is achieved with 1% labeled MS-COCO data.
  • Around 10 absolute mAP improvements over the supervised baseline are achieved with 0.5, 1, 2, and 5% labeled data.
  • The paper reports state-of-the-art semi-supervised object detection performance across COCO-standard, COCO-additional, and VOC datasets.

2 RELATED WORKS

Prior semi-supervised learning methods commonly use augmentation and consistency regularization, while semi-supervised object detection adapts these ideas to limited labeled and completely unlabeled images. Unbiased Teacher is presented as addressing limitations associated with pseudo-labeling in this setting.

  • Semi-Supervised Learning: Recent semi-supervised learning methods typically combine input augmentations or perturbations with consistency regularization.
  • Semi-Supervised Object Detection: Unbiased Teacher uses EMA to combat class imbalance and detrimental pseudo-label effects in object detection.
  • Semi-Supervised Object Detection: The model and labeled-only baseline are compared through validation losses for RPN and ROIhead classification and regression under 1% and 5% labeled-data conditions.
  • Semi-Supervised Object Detection: Semi-supervised object detection methods have explored combinations of labeled, weakly-labeled, and unlabeled data, alongside fully supervised detection.
  • Semi-Supervised Object Detection: The paper follows a standard SSL setting with a small labeled image set and a separate completely unlabeled image set.
  • Semi-Supervised Object Detection: CSD enforces prediction consistency between an image and its flipped version, while STAC generates pseudo-labels once after pretraining on limited labeled data.

3 UNBIASED TEACHER

Unbiased Teacher addresses semi-supervised object detection through staged Teacher-Student mutual learning and class-balanced pseudo-labeling. A gradually updated Teacher generates pseudo-labels, while EMA refinement and Focal loss target instability and class-imbalance bias.

  • Overview: Unbiased Teacher first initializes the detector on labeled data, then duplicates it into Teacher and Student models for mutual learning.The Burn-In stage precedes Teacher-Student Mutual Learning.
  • Teacher-Student Mutual Learning: The Teacher generates pseudo-labels from weakly augmented inputs, while the Student learns from strongly augmented inputs.The Teacher is fixed during Student Learning, and only Student weights are updated by back-propagation after pseudo-label generation.
  • Teacher-Student Mutual Learning: Class-wise NMS removes repetitive box predictions, while unsupervised regression losses are omitted because confidence scores indicate category confidence rather than localization quality.A confidence threshold filters low-confidence predicted bounding boxes before pseudo-label use.
  • Teacher-Student Mutual Learning: EMA gradually transfers Student knowledge to the Teacher, producing a slowly progressing model that acts as an ensemble across training iterations.This stabilizes pseudo-labels and helps address pseudo-labeling bias in semi-supervised object detection.
  • Bias in Pseudo-Label: Class imbalance and foreground-background imbalance impede direct transfer of image-classification pseudo-labeling methods to object detection.The paper identifies dominant classes and background instances as sources of pseudo-labeling bias.
  • Bias in Pseudo-Label: Focal loss replaces standard cross-entropy for ROI classification, emphasizing lower-confidence hard samples rather than easier examples from dominant classes.The framework also uses pseudo-labels as explicit supervision for both RPN and ROIhead to alleviate overfitting.

4 EXPERIMENTS

Experiments evaluate Unbiased Teacher across COCO and VOC settings, showing gains from evolving pseudo-label generation and class-imbalance mitigation. Ablations indicate that EMA and Focal loss improve pseudo-label balance and detection performance.

  • Experimental settings: Unbiased Teacher is evaluated on COCO-standard, COCO-additional, and VOC settings with limited or additional unlabeled data.COCO-standard samples 0.5–10% labeled training data, while COCO-additional and VOC test whether unlabeled data improves fully supervised training.
  • COCO-standard results: 20.75% mAP with 1% labeled COCO data exceeds STAC at 2%, CSD at 5%, and the supervised baseline at 5%.The method also shows around 10 absolute mAP improvements over the supervised method with less than 5% labeled data.
  • Pseudo-label quality: Teacher-Student Mutual Learning gradually evolves the Teacher, enabling more accurate pseudo-labels than a frozen pseudo-label generation model.Figure 4 examines pseudo-box accuracy, mIoU, and box counts after the 2k-iteration Burn-In stage.
  • Ablation study: EMA and Focal loss address pseudo-label class imbalance, with balanced pseudo-labels benefiting minority-class predictions.The ablations compare EMA against standard training and Focal loss against cross-entropy.
  • COCO-additional and VOC results: 1.10 absolute AP improvement is obtained on COCO-additional, while VOC07 labeled with VOC12 unlabeled yields 6.56 absolute mAP improvement.With COCO20cls as additional unlabeled data, Unbiased Teacher achieves 8.21 absolute mAP improvement.
  • Ablation study: With EMA, the Teacher is detached from Student optimization, producing smoother learning and reducing the effect of noisy pseudo-labels.The Teacher functions as a temporal ensemble of Student models across training steps.

5 CONCLUSION

The paper revisits semi-supervised object detection and identifies overfitting and class imbalance as major issues in low-labeled scenarios. Unbiased Teacher jointly trains a Teacher and Student and achieves satisfactory performance across multiple datasets.

  • Conclusion: Low-labeled object detection involves overfitting and class imbalance that affect pseudo-labeling.The paper frames these as two major issues addressed by its method.
  • Conclusion: Unbiased Teacher uses a unified Teacher-Student framework in which both models jointly learn to improve each other.The experiments report that the framework prevents pseudo-labeling bias and overfitting issues.
  • Conclusion: Unbiased Teacher achieves satisfactory performance across multiple semi-supervised object detection datasets.The conclusion summarizes the method's reported experimental outcome.

A.1 EMA ON IMBALANCED PSEUDO-LABELING ISSUE

EMA mitigates the later-training bias of pseudo-label distributions, although imbalance remains and motivates adding Focal loss.

  • EMA on imbalanced pseudo-labeling: At 30k iterations, Teacher models with and without EMA produce pseudo-label distributions close to the ground-truth distribution.Both models initially have small KL divergence from the ground-truth labels.
  • EMA on imbalanced pseudo-labeling: Without EMA, pseudo-labels become biased toward specific classes later in training, whereas EMA produces less imbalanced distributions.The Student's training on Teacher-generated pseudo-labels contributes to the divergence between the two settings.
  • EMA on imbalanced pseudo-labeling: EMA does not eliminate the balance issue, so Focal loss is added to further mitigate pseudo-label imbalance.The paper presents Focal loss as a complementary intervention.

A.2 ADDITIONAL ABLATION STUDY

The appendix extends the main-paper ablations to additional training and pseudo-labeling choices.

  • Additional ablation study: Additional ablations examine the Burn-In stage, pseudo-labeling threshold, EMA rates, and unsupervised loss weights.These studies are provided in the Appendix beyond the main ablations.

A.2.1 EFFECT OF BURN-IN STAGE

The Burn-In stage improves early pseudo-box accuracy, producing higher early training accuracy and better converged results.

  • A.2.1 EFFECT OF BURN-IN STAGE: Burn-In produces more accurate pseudo-boxes during the early training stage.The improved pseudo-box quality leads to higher accuracy early in training.
  • A.2.1 EFFECT OF BURN-IN STAGE: Models using Burn-In achieve better results after convergence.The comparison reports higher converged performance than training without Burn-In.

A.2.2 EFFECT OF PSEUDO-LABELING THRESHOLD

Confidence thresholding filters low-confidence predicted boxes, while the threshold value trades off pseudo-label quantity and quality. Burn-In also improves early pseudo-box generation and accuracy.

  • A.2.2 EFFECT OF PSEUDO-LABELING THRESHOLD: Confidence thresholding filters low-confidence predicted bounding boxes that are more likely to be false positives.The paper evaluates predicted-box accuracy before and after applying this filtering step.
  • A.2.2 EFFECT OF PSEUDO-LABELING THRESHOLD: Burn-In derives more accurate pseudo-boxes and achieves higher accuracy during early training.It also improves the final result after convergence compared with omitting Burn-In.
  • A.2.2 EFFECT OF PSEUDO-LABELING THRESHOLD: Confidence thresholding significantly improves pseudo-label quality.Accuracy is measured by comparing predicted labels with ground-truth labels before and after thresholding.
  • A.2.2 EFFECT OF PSEUDO-LABELING THRESHOLD: Lowering δ increases generated pseudo-boxes, but δ = 0.6 produces too many likely false positives while δ = 0.9 produces too few labels.Both excessively low and excessively high thresholds reduce model performance.

A.2.3 EFFECT OF EMA RATES

The Teacher model is sensitive to the EMA rate: smaller rates are less stable, while larger rates reduce variance but can slow adaptation. The experiments use α = 0.9996.

  • A.2.3 EFFECT OF EMA RATES: α = 0.5 gives the Teacher lower AP and higher variance because Student updates contribute more at each iteration.The unstable curve is attributed to detrimental effects from noisy pseudo-labels.
  • A.2.3 EFFECT OF EMA RATES: As α increases to 0.99, the Teacher gradually improves across training iterations.A larger EMA rate stabilizes and improves the learning curve relative to α = 0.5.
  • A.2.3 EFFECT OF EMA RATES: α = 0.9999 yields the lowest variance but makes the Teacher improve overly slowly.The paper therefore uses α = 0.9996 for all experiments.
  • A.2.3 EFFECT OF EMA RATES: The reported AP metric is AP50:95, following prior work.The Teacher-model analysis breaks AP down from AP50 to AP95.
Loading 2102.09480v1…