Pith. sign in

REVIEW 3 major objections 6 minor 79 references

De-Simplifying Pseudo Labels to Enhancing Domain Adaptive Object Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Simple-label bias, not label quality, is what stalls self-training detectors; DeSimPL removes it by fusing pseudo labels, adding adversarial noise, and reweighting loss.

desk verdict A competent pseudo-label update recipe with modest but consistent gains; the 'simple-label bias' diagnosis is correlational and the SOTA claim overreaches on Sim10k. read the letter →

arxiv 2507.00608 v1 pith:6CIHV75E submitted 2025-07-01 cs.CV

Zehua Fu , Chenguang Liu , Yuyu Chen , Jiaqi Zhou , Qingjie Liu , Yunhong Wang This is my paper · ORCID
classification cs.CV
keywords unsuperviseddomainadaptationobjectdetectionself-labelingpseudolabelssimple-labelbiasweightedboxfusionadversarialaugmentationadaptiveloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish why self-labeling detectors lag behind domain-alignment detectors in unsupervised domain adaptation for object detection, and it names a cause: simple-label bias, the persistent dominance of easy, low-loss true-positive pseudo labels during training. The authors argue that simply updating pseudo labels to be more accurate does not help, since SimROD with online updates reaches 46.1% mAP versus 46.5% without on VOC to Comic while the proportion of simple samples stays high. They propose DeSimPL, which reduces that proportion through an instance-level memory bank updated by weighted box fusion, FGSM-based adversarial augmentation, and an adaptive weighted loss. On the four tested benchmarks, DeSimPL lifts mAP over SimROD by 4.6, 5.1, 1.9, and 1.2 points respectively, reaching comparable or better numbers than alignment-based baselines on multiple benchmarks. If correct, the result reframes pseudo-label training as a difficulty-distribution problem rather than only a label-quality problem.

What carries the argument

The load-bearing mechanism is the simple-label bias and its countermeasure. Simple samples are defined as true-positive pseudo labels with loss at most 0.3; they dominate training yet contribute little new gradient. DeSimPL's three components are an instance-level memory bank that stores historical pseudo labels and fuses them with current low-threshold predictions through Weighted Box Fusion, preventing pseudo labels from overfitting; FGSM perturbation of DomainMix-augmented images, which creates harder inputs; and an adaptive weighted loss that scales the localization loss by the pseudo-label confidence when confidence is below 0.3. Together these keep the pseudo-label diet hard enough that each update pushes the model further.

What would settle it

Keep all three DeSimPL components identical and, in one arm, forcibly subsample or down-weight the true-positive pseudo labels with loss below 0.3 while leaving everything else fixed; if target mAP does not improve over the unmodified run, simple-label bias is not the operative cause.

Watch

Extended reading notes

Core claim

The central claim is that what limits a self-labeling detector is not the accuracy of its pseudo labels but the high proportion of simple samples, defined operationally as true-positive pseudo labels whose loss stays below roughly 0.3 throughout training. The paper shows that SimROD with online pseudo-label updates produces better labels but no better target mAP than SimROD without them, while the simple-sample proportion remains high. DeSimPL attacks this bias directly: the instance-level memory bank fuses current predictions with stored pseudo labels through Weighted Box Fusion, FGSM adversarial examples increase the presence of hard samples, and an adaptive weighted loss down-weights low-confidence localization losses to suppress late-training false positives. The result is a drop in the simple-sample proportion, faster convergence, and gains of 4.6 mAP on VOC to Comic, 5.1 on VOC to Clipart, 1.9 on Sim10k to Cityscapes, and 1.2 on KITTI to Cityscapes over SimROD.

Load-bearing premise

The load-bearing premise is that simple-label bias is a cause rather than a symptom of stalled adaptation, since DeSimPL changes pseudo-label fusion, input augmentation, and loss weighting at once, so the measured gains could come from those changes even if the easy-sample share is irrelevant.

Editorial extensions

If this is right

  • Online pseudo-label updating is not automatically beneficial: improving label quality alone left SimROD at 46.1% versus 46.5% mAP, so practitioners should monitor the share of low-loss true-positive pseudo labels as a learning signal.
  • The simple-sample proportion can serve as a training diagnostic; keeping it from staying high is tied to continued mAP growth and faster convergence.
  • Fusing current predictions into a memory bank with WBF gives finer-grained pseudo-label refinement than direct replacement or MEV-C triplet voting, because it averages coordinates and confidence across matched boxes.
  • Small FGSM perturbations applied after DomainMix, together with confidence-weighted localization loss, contribute to the gains and are usable in other self-labeling pipelines.
  • Self-labeling detectors can reach or exceed strong domain-alignment results without domain discriminators, as shown by the comparisons with D-adapt and SIGMA++ on the tested benchmarks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that simple-label bias is likely a general failure mode of iterative self-training beyond object detection; plotting the low-loss true-positive share during training in other semi-supervised tasks would show whether the same plateau appears there.
  • A natural testable extension is to reweight or drop low-loss pseudo-labels directly while keeping the memory bank and augmentation fixed, since most of the mAP gain surviving would identify de-simplification as the active ingredient rather than fused labels or adversarial noise.
  • The results suggest framing domain adaptation as a difficulty-scheduling problem, which invites combining these mechanisms with feature alignment rather than treating the two paradigms as competitors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper addresses unsupervised domain adaptation for object detection, arguing that self-labeling detectors stagnate because of 'simple-label bias'—an overabundance of easy pseudo-labeled samples during training—and proposing DeSimPL to mitigate it. DeSimPL has three components: an instance-level memory bank updated with Weighted Box Fusion, FGSM adversarial augmentation after DomainMix, and an adaptive weighted localization loss. Experiments on VOC→Comic, VOC→Clipart, Sim10k→Cityscapes, and KITTI→Cityscapes report gains over SimROD, with ablations for each component and several hyperparameters, plus analyses of pseudo-label quality, simple-label proportion, and per-city/foggy robustness.

Significance. If the causal diagnosis is accepted, the paper identifies a general failure mode of pseudo-label training and a reusable remedy, which would be valuable. The work has clear methodological anchors: explicit loss definitions (Eqs. 1–3), component-wise ablations (Table IX), benchmark comparisons (Tables II–V), and analyses of pseudo-label quality (Figures 1, 3, 6) and robustness (Tables X–XI). These are genuine strengths. However, the central causal claim currently rests on correlational evidence, and the 'state-of-the-art' wording is not supported on Sim10k→Cityscapes. The significance is therefore conditional on strengthening the causal identification and recalibrating the claims.

major comments (3)
  1. [Section III.B and Figure 3] The paper's central claim—that the high proportion of simple pseudo-labels is a cause, not a consequence, of stalled adaptation—is supported only by correlation. Across SimROD, SimROD w. online update, and DeSimPL, a lower measured simple-sample proportion co-occurs with higher mAP, but DeSimPL simultaneously changes the pseudo-label values (WBF memory bank), the input distribution (FGSM augmentation), and the loss weighting (adaptive weighted loss). No experiment varies the simple-sample proportion while holding these components fixed. The confound is especially sharp for AWL: Equation (3) multiplies low-confidence localization losses by c ≤ 1, which directly lowers their loss values; if the Figure 3 criterion (loss ≤ 0.3) is computed on the weighted loss, AWL mechanically reduces the measured simple proportion. Please add a controlled intervention (e.g., re-weighting or re-sampling to manipulate the simple proportion while fixing the other components) or weaken the causal claim to a correlated training dynamic. As written, the explanation that 'de-simplifying' is responsible for the gains is not identified.
  2. [Section IV.C, Table IV] The Introduction's claim of 'state-of-the-art results in four domain adaptive object detection benchmarks' is contradicted by Table IV: on Sim10k→Cityscapes, SIGMA++ reaches 57.7 AP and DeSimPL reaches 55.3 AP, and the text itself says DeSimPL 'outperforms most other methods' rather than all. The wording should be corrected to state which benchmarks are SOTA and which are competitive. Relatedly, Table IX reports 58.7 (Ours) and 56.8 (SimROD) for the same Sim10k→Cityscapes setting while Table IV reports 55.3 (Ours) and 53.5 (SimROD); the teacher-model versus student-model reporting must be reconciled or explicitly labeled.
  3. [Section IV.A and Section IV.D] Hyperparameter choices appear to be selected directly on the target test benchmarks. FGSM epsilon is ablated on VOC→Clipart (Table VII), the pseudo-label confidence threshold is ablated on VOC→Comic (Table VIII), and the AWL threshold tau is fixed at 0.3. No validation split or selection protocol is described. This makes the reported gains optimistic and makes the method hard to transfer. Please report a validation protocol (e.g., a held-out subset of the target training set or a source-only validation set) and state whether the final test numbers correspond to the best hyperparameter configuration.
minor comments (6)
  1. [Section IV.C] The heading 'Dissilimar domains' should be 'Dissimilar domains'.
  2. [Section IV.A and Table I] There are several typos: 'euqals' should be 'equals', 'KTIII→Cityscapes' should be 'KITTI→Cityscapes', and 'Catagories' in Table I should be 'Categories'.
  3. [Related Work and References] The text refers to 'The SoftTeacher approach [47]', but reference [47] is Instant-Teaching; SoftTeacher appears as reference [49]. Please correct the citation.
  4. [Section III.B and Figure 3] The definition of 'simple samples' as samples with loss ≤ 0.3 should specify whether the loss is the weighted localization loss of Eq. (3) or the unweighted total loss; this is important for interpreting Figure 3.
  5. [Equation (1)] Equation (1) uses y inside the gradient without defining it in that equation; clarify that y denotes the pseudo-label (box and class) used for the adversarial perturbation.
  6. [Abstract] The phrase 'adversarial samples are introduced during training to enhance the proportion' should say 'increase the proportion of hard samples', since the intended effect is to reduce the simple-label proportion.

Circularity Check

0 steps flagged · score 1.0 of 10

No step in the paper's derivation is circular by construction; the simple-label-bias diagnosis is a causal-identification concern, not a self-referential prediction.

full rationale

The paper's loss and update rules (Eqs. 1-3) are explicit and non-self-referential: FGSM perturbation, WBF fusion, and the adaptive weighted loss are all concrete mechanisms whose effects on mAP are evaluated on external benchmarks (VOC to Comic/Clipart, Sim10k to Cityscapes, KITTI to Cityscapes). The reported gains are empirical comparisons against published baselines, not quantities derived from the simple-label metric. There is no load-bearing self-citation: the cited ST3D, SimROD, SoftTeacher, and WBF works are external, and no uniqueness theorem is invoked. The central causal claim that 'simple-label bias' causes stalled adaptation is supported only by co-variation in Figure 3, where the three DeSimPL components change pseudo-label fusion, input augmentation, and loss weighting simultaneously, so the improvement cannot be uniquely attributed to de-simplifying. This is a causal-identification limitation, not a circular derivation. One potential confound is that Eq. (3) reweights the localization loss for low-confidence boxes, and Figure 3 defines 'simple samples' by loss value at most 0.3; if the Figure 3 metric uses the weighted loss, the measurement is not independent of the intervention. However, the paper does not state that the metric uses the weighted loss, and the observed direction of the proportion change is not the mechanical effect of the reweighting, so I do not treat this as a demonstrated circular step.

Assumptions & free parameters 8 free parameters · 6 assumptions · 1 invented entities

The paper's contributions are empirical; the central causal claim depends on an unproven premise (simple-label bias as cause), and the method relies on several hand-tuned thresholds and standard-but-assumed techniques (WBF, FGSM, confidence reweighting). No physical entities are introduced; the conceptual entity 'simple-label bias' has no independent falsifiable handle outside this paper's own pipeline.

free parameters (8)
  • Simple-sample loss threshold = 0.3
    Defines which pseudo-labeled samples count as 'simple' in Figure 3 and in the central diagnostic; chosen by hand, no sensitivity analysis reported.
  • AWL confidence threshold tau = 0.3
    Sets the cutoff below which localization loss is weighted by confidence (Eq. 3); grounded in Figure 6 but hand-selected.
  • FGSM epsilon = 0.01
    Perturbation strength for adversarial augmentation; chosen by ablation on VOC to Clipart (Table VII) among 0, 0.01, 0.05, 0.1.
  • WBF update prediction confidence threshold = 0.05
    Filters current-model predictions before fusion with the memory bank; chosen by ablation on VOC to Comic (Table VIII).
  • Initial pseudo-label confidence threshold = 0.6
    Used when generating the initial target pseudo-labels from the source-pretrained teacher; hand-set, no ablation reported.
  • WBF IoU matching threshold = 0.5
    Groups boxes for weighted fusion; standard but still a manual choice.
  • Pseudo-label update interval = 1 or 10 epochs depending on benchmark
    Per-setting schedule in Section IV.A; influences how often the memory bank refreshes.
  • Student model confidence threshold = 0.2 (VOC to Comic), 0.3 (others)
    Filters pseudo-labels when training the student model; hand-set per benchmark.
assumptions (6)
  • domain assumption The conditional label distribution is shared across domains: p_S(y|x)=p_T(y|x), while input distributions differ.
    Standard UDA assumption stated in Section III.A; all benchmarks rely on it.
  • domain assumption Pseudo-labels from a high-confidence source-trained teacher provide a valid supervisory signal for target training.
    Inherited from SimROD (Section III.C); the entire self-labeling framework depends on teacher pseudo-labels being useful.
  • ad hoc to paper Weighted Box Fusion of current predictions with the memory bank produces more accurate pseudo-labels than direct replacement or MEV-C.
    Chosen in Section III.C.1 and supported only by the ILMB ablation rows of Table IX; the mechanism is not isolated.
  • ad hoc to paper FGSM adversarial perturbation increases the proportion of hard samples and improves target-domain generalization without destroying the training signal.
    Invoked in Section III.C.2; the epsilon sweep shows sensitivity but no measurement of the hard-sample pathway.
  • ad hoc to paper Low-confidence pseudo-labels are mostly false positives, so downweighting their localization loss by confidence is beneficial.
    Based on Figure 6, a single diagnostic across three epochs; the threshold 0.3 is hand-set.
  • ad hoc to paper The high proportion of simple pseudo-labels is a cause, not a consequence, of stagnating performance.
    The paper's key premise; evidence is correlational (Figure 3, Section III.B) and confounded with the components of DeSimPL.
invented entities (1)
  • simple-label bias
    purpose: A hypothesized systematic over-representation of easy target pseudo-labels (loss below 0.3) in self-labeling UDA training, proposed as the central cause of stalled performance.
    The paper supports it only with an in-pipeline correlation (Figure 3) and the success of DeSimPL, which was designed to reduce the bias. No external dataset or controlled manipulation isolates the bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of De-Simplifying Pseudo Labels to Enhancing Domain Adaptive Object Detection." pith.science (2026). https://pith.science/paper/6CIHV75E

@misc{pith2026250700608,
  author       = {Pith},
  title        = {Pith review of: De-Simplifying Pseudo Labels to Enhancing Domain Adaptive Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6CIHV75E}},
  note         = {Machine review of arXiv:2507.00608}
}
read the original abstract

Despite its significant success, object detection in traffic and transportation scenarios requires time-consuming and laborious efforts in acquiring high-quality labeled data. Therefore, Unsupervised Domain Adaptation (UDA) for object detection has recently gained increasing research attention. UDA for object detection has been dominated by domain alignment methods, which achieve top performance. Recently, self-labeling methods have gained popularity due to their simplicity and efficiency. In this paper, we investigate the limitations that prevent self-labeling detectors from achieving commensurate performance with domain alignment methods. Specifically, we identify the high proportion of simple samples during training, i.e., the simple-label bias, as the central cause. We propose a novel approach called De-Simplifying Pseudo Labels (DeSimPL) to mitigate the issue. DeSimPL utilizes an instance-level memory bank to implement an innovative pseudo label updating strategy. Then, adversarial samples are introduced during training to enhance the proportion. Furthermore, we propose an adaptive weighted loss to avoid the model suffering from an abundance of false positive pseudo labels in the late training period. Experimental results demonstrate that DeSimPL effectively reduces the proportion of simple samples during training, leading to a significant performance improvement for self-labeling detectors. Extensive experiments conducted on four benchmarks validate our analysis and conclusions.

Figures

Figures reproduced from arXiv: 2507.00608 by the authors.

Figure 1
Figure 1. In the self-labeling paradigm, online updating is an [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our DeSimPL comprises three components: an online update pseudo label strategy based on the instance-level memory [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The proportion variations of simple samples (i.e., [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Paradigm of self-training in domain adaptive object [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: The relationship between the rate of false positive [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparisons under the setting Sim10k [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparisons under the setting Pascal VOC [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Visualization of adversarial examples. complex scenes, emphasizing the method’s robustness under challenging domain shifts. Adversarial data augmentation [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Visualization of pseudo labels during training. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 10
Figure 10. Figure 10: Visualization of pseudo labels during training. (cont.) [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 58 canonical work pages

  1. [1]

    Faster r-cnn: Towards real-time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,”Advances in neural information processing systems, vol. 28, 2015

    2015

  2. [2]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779– 788. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    work page 2016

  3. [3]

    Ssd: Single shot multibox detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” inComputer Vision– ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. Springer, 2016, pp. 21–37

    2016

  4. [4]

    Spatially invariant unsupervised object detection with convolutional neural networks,

    E. Crawford and J. Pineau, “Spatially invariant unsupervised object detection with convolutional neural networks,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 01, pp. 3412–3420, Jul. 2019. [Online]. Available: https://ojs.aaai.org/index. php/AAAI/article/view/4216

    work page 2019

  5. [5]

    Category dictionary guided unsupervised domain adaptation for object detection,

    S. Li, J. Huang, X.-S. Hua, and L. Zhang, “Category dictionary guided unsupervised domain adaptation for object detection,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 3, pp. 1949–1957, May 2021. [Online]. Available: https://ojs.aaai.org/index. php/AAAI/article/view/16290

    work page 1949

  6. [6]

    Shape-adaptive selection and measurement for oriented object detection,

    L. Hou, K. Lu, J. Xue, and Y . Li, “Shape-adaptive selection and measurement for oriented object detection,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 1, pp. 923–932, Jun

    work page

  7. [7]

    Weakly-supervised camouflaged object detection with scribble annotations,

    R. He, Q. Dong, J. Lin, and R. W.H. Lau, “Weakly-supervised camouflaged object detection with scribble annotations,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, pp. 781–789, Jun. 2023. [Online]. Available: https://ojs.aaai.org/index.php/ AAAI/article/view/25156

    work page 2023

  8. [8]

    A review and comparative study on probabilistic object detection in autonomous driv- ing,

    D. Feng, A. Harakeh, S. L. Waslander, and K. Dietmayer, “A review and comparative study on probabilistic object detection in autonomous driv- ing,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 8, pp. 9961–9980, 2022

    work page 2022

Show all 79 references
  1. [9]

    C2FDA: coarse-to-fine domain adaptation for traffic object detection,

    H. Zhang, G. Luo, J. Li, and F. Wang, “C2FDA: coarse-to-fine domain adaptation for traffic object detection,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 8, pp. 12 633–12 647, 2022. [Online]. Available: https://doi.org/10.1109/TITS.2021.3115823

    2022

  2. [10]

    Aibm: Accurate and instant background modeling for moving object detection,

    W. J. Kim, S. Hwang, J. Lee, S. Woo, and S. Lee, “Aibm: Accurate and instant background modeling for moving object detection,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 7, pp. 9021–9036, 2022

    2022

  3. [11]

    A cross-scale and illumination invariance-based model for robust object detection in traffic surveillance scenarios,

    Y .-F. Lu, J.-W. Gao, Q. Yu, Y . Li, Y .-S. Lv, and H. Qiao, “A cross-scale and illumination invariance-based model for robust object detection in traffic surveillance scenarios,”IEEE Transactions on Intelligent Trans- portation Systems, vol. 24, no. 7, pp. 6989–6999, 2023

    2023

  4. [12]

    Tfgnet: Traffic salient object detection using a feature deep interaction and guidance fusion,

    N. Jia, Y . Sun, and X. Liu, “Tfgnet: Traffic salient object detection using a feature deep interaction and guidance fusion,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 3, pp. 3020–3030, 2024

    2024

  5. [13]

    Domain adaptive faster r-cnn for object detection in the wild,

    Y . Chen, W. Li, C. Sakaridis, D. Dai, and L. Van Gool, “Domain adaptive faster r-cnn for object detection in the wild,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3339– 3348

    2018

  6. [14]

    Strong-weak distri- bution alignment for adaptive object detection,

    K. Saito, Y . Ushiku, T. Harada, and K. Saenko, “Strong-weak distri- bution alignment for adaptive object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 6956–6965

    2019

  7. [15]

    Every pixel matters: Center-aware feature alignment for domain adaptive object detector,

    C.-C. Hsu, Y .-H. Tsai, Y .-Y . Lin, and M.-H. Yang, “Every pixel matters: Center-aware feature alignment for domain adaptive object detector,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16. Springer, 2020, pp. 733–748

    2020

  8. [16]

    Prior-based domain adaptive object detection for hazy and rainy conditions,

    V . A. Sindagi, P. Oza, R. Yasarla, and V . M. Patel, “Prior-based domain adaptive object detection for hazy and rainy conditions,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16. Springer, 2020, pp. 763–780

    2020

  9. [17]

    Mega- cda: Memory guided attention for category-aware unsupervised domain adaptive object detection,

    V . Vs, V . Gupta, P. Oza, V . A. Sindagi, and V . M. Patel, “Mega- cda: Memory guided attention for category-aware unsupervised domain adaptive object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4516–4526

    2021

  10. [18]

    Multi-adversarial faster-rcnn for unrestricted object detection,

    Z. He and L. Zhang, “Multi-adversarial faster-rcnn for unrestricted object detection,” inProceedings of the IEEE/CVF International Con- ference on Computer Vision, 2019, pp. 6668–6677

    2019

  11. [19]

    Progressive domain adaptation for object detection,

    H.-K. Hsu, C.-H. Yao, Y .-H. Tsai, W.-C. Hung, H.-Y . Tseng, M. Singh, and M.-H. Yang, “Progressive domain adaptation for object detection,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2020, pp. 749–757

    2020

  12. [20]

    Diversify and match: A domain adaptive representation learning paradigm for object detection,

    T. Kim, M. Jeong, S. Kim, S. Choi, and C. Kim, “Diversify and match: A domain adaptive representation learning paradigm for object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 12 456–12 465

    2019

  13. [21]

    Cross-domain object detection through coarse-to-fine feature adaptation,

    Y . Zheng, D. Huang, S. Liu, and Y . Wang, “Cross-domain object detection through coarse-to-fine feature adaptation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 13 766–13 775

    2020

  14. [22]

    Adapting object detectors via selective cross-domain alignment,

    X. Zhu, J. Pang, C. Yang, J. Shi, and D. Lin, “Adapting object detectors via selective cross-domain alignment,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 687– 696

    2019

  15. [23]

    Decoupled adaptation for cross-domain object detection,

    J. Jiang, B. Chen, J. Wang, and M. Long, “Decoupled adaptation for cross-domain object detection,” inThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022

    2022

  16. [24]

    Sigma: Semantic-complete graph matching for domain adaptive object detection,

    W. Li, X. Liu, and Y . Yuan, “Sigma: Semantic-complete graph matching for domain adaptive object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5291–5300

    2022

  17. [25]

    Simrod: A simple adaptation method for robust object detection,

    R. Ramamonjison, A. Banitalebi-Dehkordi, X. Kang, X. Bai, and Y . Zhang, “Simrod: A simple adaptation method for robust object detection,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3570–3579

    2021

  18. [26]

    Automatic adaptation of object detectors to new domains using self-training,

    A. RoyChowdhury, P. Chakrabarty, A. Singh, S. Jin, H. Jiang, L. Cao, and E. Learned-Miller, “Automatic adaptation of object detectors to new domains using self-training,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 780– 790

    2019

  19. [27]

    A robust learning approach to domain adaptive object detection,

    M. Khodabandeh, A. Vahdat, M. Ranjbar, and W. G. Macready, “A robust learning approach to domain adaptive object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 480–490

    2019

  20. [28]

    Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection,

    S. Kim, J. Choi, T. Kim, and C. Kim, “Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection,” inProceedings of the IEEE/CVF International Con- ference on Computer Vision, 2019, pp. 6092–6101

    2019

  21. [29]

    Collaborative training between region proposal localization and classification for domain adaptive object detection,

    G. Zhao, G. Li, R. Xu, and L. Lin, “Collaborative training between region proposal localization and classification for domain adaptive object detection,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII

    2020

  22. [30]

    Unsupervised domain adaptation for object detection via cross-domain semi-supervised learning,

    F. Yu, D. Wang, Y . Chen, N. Karianakis, T. Shen, P. Yu, D. Lymberopou- los, S. Lu, W. Shi, and X. Chen, “Unsupervised domain adaptation for object detection via cross-domain semi-supervised learning,”arXiv preprint arXiv:1911.07158, 2019

    1911 arXiv

  23. [31]

    Springer, 2020, pp. 86–102

    2020

  24. [32]

    Robust object detection via instance-level temporal cycle confusion,

    X. Wang, T. E. Huang, B. Liu, F. Yu, X. Wang, J. E. Gonzalez, and T. Darrell, “Robust object detection via instance-level temporal cycle confusion,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9143–9152

    2021

  25. [33]

    Sstn: Self-supervised domain adapta- tion thermal object detection for autonomous driving,

    F. Munir, S. Azam, and M. Jeon, “Sstn: Self-supervised domain adapta- tion thermal object detection for autonomous driving,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 206–213

    2021

  26. [34]

    Sigma++: Improved semantic-complete graph matching for domain adaptive object detection,

    W. Li, X. Liu, and Y . Yuan, “Sigma++: Improved semantic-complete graph matching for domain adaptive object detection,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 2023

    2023

  27. [35]

    Cross-domain weakly-supervised object detection through progressive domain adap- tation,

    N. Inoue, R. Furuta, T. Yamasaki, and K. Aizawa, “Cross-domain weakly-supervised object detection through progressive domain adap- tation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 5001–5009

    2018

  28. [36]

    Deep domain adaptation for pavement crack detection,

    H. Liu, C. Yang, A. Li, S. Huang, X. Feng, Z. Ruan, and Y . Ge, “Deep domain adaptation for pavement crack detection,”IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 2, pp. 1669–1681, 2023

    2023

  29. [37]

    Stepwise domain adaptation (SDA) for object detection in autonomous vehicles using an adaptive centernet,

    G. Li, Z. Ji, and X. Qu, “Stepwise domain adaptation (SDA) for object detection in autonomous vehicles using an adaptive centernet,”IEEE Trans. Intell. Transp. Syst., vol. 23, no. 10, pp. 17 729–17 743, 2022. [Online]. Available: https://doi.org/10.1109/TITS.2022.3164407

    2022

  30. [38]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 2980–2988

    2017

  31. [39]

    Progressive critical region transfer for cross-domain visual object detection,

    X. Wang, P. Jiang, Y . Li, M. Hu, M. Gao, D. Cao, and R. Ding, “Progressive critical region transfer for cross-domain visual object detection,”IEEE Transactions on Intelligent Transportation Systems, pp. 1–15, 2024

    2024

  32. [40]

    Exploring categorical regularization for domain adaptive object detection,

    C.-D. Xu, X.-R. Zhao, X. Jin, and X.-S. Wei, “Exploring categorical regularization for domain adaptive object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 11 724–11 733

    2020

  33. [41]

    Seeking similarities over differences: Similarity-based do- main alignment for adaptive object detection,

    F. Rezaeianaran, R. Shetty, R. Aljundi, D. O. Reino, S. Zhang, and B. Schiele, “Seeking similarities over differences: Similarity-based do- main alignment for adaptive object detection,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9204–...

    2021

  34. [42]

    Cycle-consistent domain adaptive faster rcnn,

    D. Zhang, J. Li, L. Xiong, L. Lin, M. Ye, and S. Yang, “Cycle-consistent domain adaptive faster rcnn,”IEEE Access, vol. 7, pp. 123 903–123 911, 2019

    2019

  35. [43]

    Adaptive object detection with dual multi-label prediction,

    Z. Zhao, Y . Guo, H. Shen, and J. Ye, “Adaptive object detection with dual multi-label prediction,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVIII 16. Springer, 2020, pp. 54–69

    2020

  36. [44]

    Domain adaptation for object detection via style consistency,

    A. L. Rodriguez and K. Mikolajczyk, “Domain adaptation for object detection via style consistency,”arXiv preprint arXiv:1911.10033, 2019

    1911 arXiv

  37. [45]

    Harmonizing transferability and discriminability for adapting object detectors,

    C. Chen, Z. Zheng, X. Ding, Y . Huang, and Q. Dou, “Harmonizing transferability and discriminability for adapting object detectors,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 8869–8878

    2020

  38. [46]

    A simple semi-supervised learning framework for object detection,

    K. Sohn, Z. Zhang, C.-L. Li, H. Zhang, C.-Y . Lee, and T. Pfister, “A simple semi-supervised learning framework for object detection,”arXiv preprint arXiv:2005.04757, 2020

    2005 arXiv

  39. [47]

    Cycada: Cycle-consistent adversarial domain adapta- tion,

    J. Hoffman, E. Tzeng, T. Park, J.-Y . Zhu, P. Isola, K. Saenko, A. Efros, and T. Darrell, “Cycada: Cycle-consistent adversarial domain adapta- tion,” inInternational conference on machine learning. Pmlr, 2018, pp. 1989–1998

    2018

  40. [48]

    St3d: Self-training for unsupervised domain adaptation on 3d object detection,

    J. Yang, S. Shi, Z. Wang, H. Li, and X. Qi, “St3d: Self-training for unsupervised domain adaptation on 3d object detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 10 368–10 378

    2021

  41. [49]

    Instant-teaching: An end- to-end semi-supervised object detection framework,

    Q. Zhou, C. Yu, Z. Wang, Q. Qian, and H. Li, “Instant-teaching: An end- to-end semi-supervised object detection framework,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4081–4090

    2021

  42. [50]

    Weighted boxes fusion: Ensembling boxes from different object detection models,

    R. Solovyev, W. Wang, and T. Gabruseva, “Weighted boxes fusion: Ensembling boxes from different object detection models,”Image and Vision Computing, vol. 107, p. 104117, 2021

    2021

  43. [51]

    End-to-end semi-supervised object detection with soft teacher,

    M. Xu, Z. Zhang, H. Hu, J. Wang, L. Wang, F. Wei, X. Bai, and Z. Liu, “End-to-end semi-supervised object detection with soft teacher,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3060–3069

    2021

  44. [52]

    Soft-nms–improving object detection with one line of code,

    N. Bodla, B. Singh, R. Chellappa, and L. S. Davis, “Soft-nms–improving object detection with one line of code,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 5561–5569

    2017

  45. [53]

    Efficient non-maximum suppression,

    A. Neubeck and L. Van Gool, “Efficient non-maximum suppression,” in 18th international conference on pattern recognition (ICPR’06), vol. 3. IEEE, 2006, pp. 850–855

    2006

  46. [54]

    ultralytics/yolov5: v1.0 - initial release,

    G. J. et al., “ultralytics/yolov5: v1.0 - initial release,” Jun. 2020. [Online]. Available: https://doi.org/10.5281/zenodo.3908560

    2020 doi

  47. [55]

    Explaining and harnessing adversarial examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,”arXiv preprint arXiv:1412.6572, 2014

    2014 arXiv

  48. [56]

    The pascal visual object classes (voc) challenge,

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisser- man, “The pascal visual object classes (voc) challenge,”International journal of computer vision, vol. 88, pp. 303–308, 2009

    2009

  49. [57]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” inComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer, 2...

    2014

  50. [58]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3213– 3223

    2016

  51. [59]

    Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks?

    M. Johnson-Roberson, C. Barto, R. Mehta, S. N. Sridhar, K. Rosaen, and R. Vasudevan, “Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks?”arXiv preprint arXiv:1610.01983, 2016

    2016 arXiv

  52. [60]

    Scl: Towards accurate domain adaptive object detection via gradient detach based stacked complementary losses,

    Z. Shen, H. Maheshwari, W. Yao, and M. Savvides, “Scl: Towards accurate domain adaptive object detection via gradient detach based stacked complementary losses,”arXiv preprint arXiv:1911.02559, 2019

    1911 arXiv

  53. [61]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in2012 IEEE conference on computer vision and pattern recognition. IEEE, 2012, pp. 3354–3361

    2012

  54. [62]

    Unbiased teacher for semi-supervised object detection,

    Y .-C. Liu, C.-Y . Ma, Z. He, C.-W. Kuo, K. Chen, P. Zhang, B. Wu, Z. Kira, and P. Vajda, “Unbiased teacher for semi-supervised object detection,”arXiv preprint arXiv:2102.09480, 2021

    2021 arXiv

  55. [63]

    Domain adaptive object detection via asymmetric tri-way faster-rcnn,

    Z. He and L. Zhang, “Domain adaptive object detection via asymmetric tri-way faster-rcnn,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIV

    2020

  56. [64]

    Springer, 2020, pp. 309–324

    2020

  57. [65]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inInternational conference on machine learning. pmlr, 2015, pp. 448–456

    2015

  58. [66]

    Adversarial discrim- inative domain adaptation,

    E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell, “Adversarial discrim- inative domain adaptation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 7167–7176

    2017

  59. [67]

    The many faces of robustness: A critical analysis of out-of-distribution generalization,

    D. Hendrycks, S. Basart, N. Mu, S. Kadavath, F. Wang, E. Dorundo, R. Desai, T. Zhu, S. Parajuli, M. Guoet al., “The many faces of robustness: A critical analysis of out-of-distribution generalization,” in Proceedings of the IEEE/CVF International Conference on Computer Vision,...

    2021

  60. [68]

    Multi-level domain adaptive learning for cross-domain detection,

    R. Xie, F. Yu, J. Wang, Y . Wang, and L. Zhang, “Multi-level domain adaptive learning for cross-domain detection,” inProceedings of the IEEE/CVF international conference on computer vision workshops, 2019, pp. 0–0

    2019

  61. [69]

    Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness,

    R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel, “Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness,”arXiv preprint arXiv:1811.12231, 2018

    2018 arXiv

  62. [70]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 2223–2232

    2017

  63. [71]

    Exploring object relation in mean teacher for cross-domain detection,

    Q. Cai, Y . Pan, C.-W. Ngo, X. Tian, L. Duan, and T. Yao, “Exploring object relation in mean teacher for cross-domain detection,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 11 457–11 466

    2019

  64. [72]

    Adapting object detectors with conditional domain normalization,

    P. Su, K. Wang, X. Zeng, S. Tang, D. Chen, D. Qiu, and X. Wang, “Adapting object detectors with conditional domain normalization,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16. Springer, 2020, pp. 403–419

    2020

  65. [73]

    Unbiased mean teacher for cross- domain object detection,

    J. Deng, W. Li, Y . Chen, and L. Duan, “Unbiased mean teacher for cross- domain object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4091–4101

    2021

  66. [74]

    Towards robust adaptive object detection under noisy annotations,

    X. Liu, W. Li, Q. Yang, B. Li, and Y . Yuan, “Towards robust adaptive object detection under noisy annotations,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. IEEE, 2022, pp. 14 187–14 196. [Online]. Availa...

    2022

  67. [75]

    Knowledge mining and transferring for domain adaptive object detection,

    K. Tian, C. Zhang, Y . Wang, S. Xiang, and C. Pan, “Knowledge mining and transferring for domain adaptive object detection,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 9133–9142

    2021

  68. [76]

    Ssal: Synergizing between self-training and adversarial learning for domain adaptive object detection,

    M. A. Munir, M. H. Khan, M. Sarfraz, and M. Ali, “Ssal: Synergizing between self-training and adversarial learning for domain adaptive object detection,”Advances in Neural Information Processing Systems, vol. 34, pp. 22 770–22 782, 2021

    2021

  69. [78]

    Novel scenes & classes: Towards adaptive open-set object detection,

    W. Li, X. Guo, and Y . Yuan, “Novel scenes & classes: Towards adaptive open-set object detection,” inIEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023. IEEE, 2023, pp. 15 734–15 744. [Online]. Available: https://doi.org/10.1109/I...

    2023

  70. [79]

    Decoupled unbiased teacher for source-free domain adaptive medical object detection,

    X. Liu, W. Li, and Y . Yuan, “Decoupled unbiased teacher for source-free domain adaptive medical object detection,”IEEE Trans. Neural Networks Learn. Syst., vol. 35, no. 6, pp. 7287–7298, 2024. [Online]. Available: https://doi.org/10.1109/TNNLS.2023.3272389

    2024

  71. [2022]

    Available: https://ojs.aaai.org/index.php/AAAI/article/ view/19975

    [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/ view/19975

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.