PaperlayerView exact revision on arXiv

Source-linked AI summary

Unbiased Mean Teacher for Cross-domain Object Detection

Jinhong Deng, Wen Li, Yuhua Chen, Lixin Duan

arXiv:2003.00707v2cs.CV

TL;DR

Cross-domain object detection is challenged by domain shift and Mean Teacher bias toward labeled source data, which can weaken target-domain guidance. UMT remedies this bias with cross-domain distillation, pixel-level source/target adaptation, and out-of-distribution sample selection, reaching state-of-the-art results on multiple benchmarks. The authors also identify static image-transformation strategies as a limitation during training.

  • Problem

    Cross-domain object detection lacks reliable target-domain guidance because Mean Teacher can become biased toward the source domain under substantial domain shift.

  • Method

    UMT combines cross-domain distillation, target-like source augmentation for student training, and out-of-distribution estimation to select suitable target samples.

  • Results

    UMT achieves state-of-the-art performance across multiple benchmarks, with reported mAPs including 44.1% on Clipart1k, 41.7% on Foggy Cityscapes, and 43.1% on Cityscapes.

  • Takeaways & Limitations

    Addressing model bias in both teacher and student models provides a basis for stronger cross-domain object detection performance across benchmark datasets.

  • Takeaways & Limitations

    The source-like and target-like image strategies are static and may become unsuitable for the teacher as training progressively shifts the model toward the target domain.

Abstract

from arXiv · show

Cross-domain object detection is challenging, because object detection model is often vulnerable to data variance, especially to the considerable domain shift between two distinctive domains. In this paper, we propose a new Unbiased Mean Teacher (UMT) model for cross-domain object detection. We reveal that there often exists a considerable model bias for the simple mean teacher (MT) model in cross-domain scenarios, and eliminate the model bias with several simple yet highly effective strategies. In particular, for the teacher model, we propose a cross-domain distillation method for MT to maximally exploit the expertise of the teacher model. Moreover, for the student model, we alleviate its bias by augmenting training samples with pixel-level adaptation. Finally, for the teaching process, we employ an out-of-distribution estimation strategy to select samples that most fit the current model to further enhance the cross-domain distillation process. By tackling the model bias issue with these strategies, our UMT model achieves mAPs of 44.1%, 58.1%, 41.7%, and 43.1% on benchmark datasets Clipart1k, Watercolor2k, Foggy Cityscapes, and Cityscapes, respectively, which outperforms the existing state-of-the-art results in notable margins. Our implementation is available at https://github.com/kinredon/umt.

1. Introduction

Cross-domain object detection must handle substantial domain shifts and the coupled classification-localization outputs of detection. UMT addresses mean-teacher model bias through cross-domain distillation, pixel-level augmentation, and adaptive sample selection, achieving state-of-the-art results across multiple benchmarks.

  • Motivation: Cross-domain detection is difficult because models often lose performance when applied beyond their training domain, while annotations across real-world conditions are costly.Autonomous driving requires robustness across weather and lighting conditions without labeled target data.
  • Motivation: Object detection is especially vulnerable to domain variance because class labels and bounding-box locations are coupled outputs.Prior responses include instance- and image-level adversarial training and graph-based consistency.
  • Mean Teacher baseline: Mean Teacher provides a starting point by distilling consistency between teacher and student predictions on perturbed unlabeled samples.The approach was originally proposed for semi-supervised learning and offers some robustness to data variance.
  • Mean Teacher bias: Large domain gaps bias Mean Teacher toward the source domain because supervision mainly comes from labeled source data, weakening target-image guidance.This bias is particularly consequential for detection because localization errors can alter feature pooling and class prediction.
  • UMT contributions: UMT counters model bias with source-like target images for the teacher, target-like source images for student training, and out-of-distribution estimation for selecting suitable teaching samples.The final UMT model reports mAPs of 44.1%, 58.1%, 41.7%, and 43.1% on Clipart1k, Watercolor2k, Foggy Cityscapes, and Cityscapes, respectively.

2. Related Works

Related work on unsupervised domain adaptation includes discrepancy minimization and adversarial domain-invariant representation learning. Cross-domain detection extends these ideas with methods addressing detection-specific domain shifts through adversarial, consistency, and graph-based techniques.

  • Unsupervised Domain Adaptation: Unsupervised domain adaptation adapts models from labeled source domains to unlabeled target domains.Prior methods minimize discrepancies such as maximum mean discrepancy or learn domain-invariant representations with adversarial domain classifiers.
  • Cross Domain Object Detection: Cross-domain object detection methods use diverse techniques, including gradient-reversal adversarial learning and other approaches for domain-invariant features.The cited methods include DA-Faster, SCDA, SWDA, SPLAT, MAF, MDAL, and CRDA.
  • Cross Domain Object Detection: The related literature targets domain shift in detection through multiple methodological families rather than a single adaptation strategy.The passage identifies adversarial learning among the techniques used in cross-domain object detection.

3. The Unbiased Mean Teacher Model

The Unbiased Mean Teacher model adapts Mean Teacher for cross-domain object detection by addressing bias in the teacher and student models and by improving sample selection during teaching. It combines cross-domain distillation, pixel-level adaptation, and jointly optimized training routes.

  • Mean Teacher foundation: Mean Teacher uses a student trained on labeled data and a teacher updated with exponential moving average weights, whose predictions are more robust and stable.The teacher therefore provides consistency-based supervision for perturbed unlabeled samples.
  • Joint training: The method trains with labeled source, source-like, target-like, and target images, combining source detection, cross-domain distillation, and additional target-like detection losses.The target domain is unlabeled, while source images provide bounding-box and class-label annotations.
  • Model bias: Cross-domain Mean Teacher can become biased toward the source domain because supervision mainly comes from labeled source samples, making teacher predictions on target images deficient.Teacher APs generally outperform on source-like samples compared with target samples across datasets.
  • Healing the teacher model bias: Cross-domain distillation feeds source-like target images to the teacher and original target images to the student, improving teacher guidance while encouraging target-domain preference.Source-like images are generated through pixel-level translation, and the teacher-side images receive small perturbations.
  • Healing the student model bias: Target-like versions of labeled source images are added to student training alongside original source samples to reduce student-model bias.The target-like images are generated with CycleGAN.
  • Teaching-process limitation: Static source-like and target-like image strategies may become unsuitable as training progresses because the model is expected to gradually fit the target domain.This motivates adapting the teaching process rather than relying only on fixed translated samples.

4. Experiments

Experiments evaluate UMT under real-to-artistic, normal-to-adverse-weather, and synthetic-to-real domain shifts using labeled source images and unlabeled target images. Across these settings, the proposed components progressively improve mean AP and achieve reported state-of-the-art results.

  • Experimental Setup: Experiments cover three domain-shift types: real-to-artistic, normal-weather-to-adverse-weather, and synthetic-to-real adaptation.The evaluation compares UMT with existing methods on benchmark datasets.
  • Experimental Setup: Unsupervised adaptation uses fully annotated source training data and unlabeled target images, with the target test set held out during training.The target domain contributes only unlabeled training images under the adopted protocol.
  • Ablation Study: Ablations define UMTS as simple MT, UMTSC as MT with cross-domain distillation, and UMTSCA as additionally using target-like images for student training.These variants isolate the contributions of the proposed components.
  • Real to Artistic Adaptation: 44.1% mAP is achieved by UMT on Clipart1k, improving over 33.6% for UMTS and 41.8% for UMTSCA.UMT also reaches the new state-of-the-art result on this dataset.
  • Adaptation in Inverse Weather: 41.7% mean AP is achieved on Foggy Cityscapes, exceeding HTCN's 39.8% and the reported oracle result.UMTSC reaches 39.2% before the student-bias and teaching-process improvements are added.
  • Synthetic-to-Real Adaptation: 43.1% AP is achieved on Cityscapes for SIM10K-to-Cityscapes adaptation using VGG16 as the backbone.The result is reported as a new state-of-the-art AP on this dataset.

5. Conclusion

The conclusion frames source-domain bias as a central issue in cross-domain object detection and presents UMT as a three-strategy response. Experiments on multiple benchmarks report that UMT surpasses existing state-of-the-art models by relatively large margins.

  • Conclusion: UMT studies cross-domain object detection through the observation that detection models can become biased toward source images.The paper presents this as a perspective for analyzing the problem.
  • Conclusion: UMT combines cross-domain distillation, pixel-level adaptation for student training, and out-of-distribution sample selection.The three strategies target teacher expertise, student model bias, and the teaching process, respectively.
  • Conclusion: Experiments on multiple benchmark datasets show that UMT surpasses existing state-of-the-art models by relatively large margins.The conclusion reports this as the overall experimental finding.

Appendices

The appendix section introduces analyses of detection errors, parameters, and qualitative results.

  • Appendices: The appendix reports detection error analysis, parameter analysis, and qualitative results.These analyses are presented as supplementary evaluations.

Appendix A. Detection Error Analysis

Detection-error analysis compares the source-only baseline, UMT, and progressively ablated UMT variants on PASCAL VOC-to-Clipart1k adaptation.

  • Appendix A. Detection Error Analysis: The analysis examines detection errors for Source Only, UMT, UMTS, UMTSC, and UMTSCA on PASCAL VOC→Clipart1k.The experimental settings remain the same as in the main paper.

A.1. Localization Errors

UMT progressively reduces localization-related detection errors compared with Source Only and simpler mean-teacher variants, with gains associated with its bias-correction strategies.

  • A.1. Localization Errors: 53.60% detection accuracy is achieved after integrating UMT’s strategies for addressing mean-teacher bias.The reported progression is 46.60%, 48.45%, 51.60%, and 53.60% across increasingly complete variants.
  • A.1. Localization Errors: 46.60% detection accuracy improves over Source Only’s 44.40%, while localization error falls from 17.65% to 14.60%.This comparison uses the simple mean-teacher version, UMTS.
  • A.1. Localization Errors: 48.45% detection accuracy with cross-domain distillation is linked mainly to reducing Mis-Classified error from 14.60% to 11.15%.The distillation strategy uses source-like samples to correct teacher-model bias.
  • A.1. Localization Errors: 51.60% detection accuracy follows the use of target-like samples to heal student-model bias.The final UMT configuration further adds out-of-distribution estimation for sample selection.

A.2. Classification Errors on Maximum Overlap Detections

Classification errors remain prominent relative to localization errors, but UMT improves the balance of correct and mis-classified detections over Source Only.

  • A.2. Classification Errors on Maximum Overlap Detections: Classification error is relatively more prominent than localization error in cross-domain object detection.The analysis evaluates detections having the highest overlap with ground truth.
  • A.2. Classification Errors on Maximum Overlap Detections: UMT increases correct-classified detections and reduces Mis-Classified detections compared with Source Only.The comparison is shown in the classification-error analysis on maximum-overlap detections.

Appendix B. Parameter Analysis

Parameter analyses examine the sensitivity of UMT to λ, γ, T, and IoU thresholds, finding robust performance across the tested settings and consistent superiority over comparison methods.

  • Appendix B. Parameter Analysis: γ and T produce consistently high detection performance across the tested parameter values, while λ is analyzed separately for sensitivity.The default settings are λ = 0.01, γ = 0.1, and T = 0.8.
  • Appendix B. Parameter Analysis: mAP drops dramatically as the IoU threshold increases because higher thresholds impose more rigorous localization requirements.The experiment compares Source Only, CRDA, and UMT on PASCAL VOC→Clipart1k adaptation.
  • Appendix B. Parameter Analysis: UMT consistently outperforms CRDA and Source Only across the tested IoU thresholds.The result suggests more robust and accurate bounding-box regression under varying evaluation thresholds.

Appendix C. Qualitative Results

Qualitative visualizations across four adaptation scenarios show UMT detections alongside baselines or ground truth, with detections displayed above a 0.6 category-score threshold.

  • Appendix C. Qualitative Results: Across Clipart1k, Watercolor2k, Foggy Cityscapes, and Cityscapes, the paper reports that UMT accurately detects objects in the images.The visualized objects include person, aeroplane, motorbike, and car examples.
  • Appendix C. Qualitative Results: Clipart1k and Watercolor2k visualizations compare Source Only, UMT, and ground truth from left to right.Both figures use detections with category scores higher than 0.6.
  • Appendix C. Qualitative Results: Foggy Cityscapes and Cityscapes visualizations show UMT detections for Cityscapes→Foggy Cityscapes and SIM10K→Cityscapes scenes.Displayed detections have category scores higher than 0.6.
Loading 2003.00707v2…