Pith. sign in

REVIEW 4 major objections 6 minor 47 references

Efficient Test-time Adaptive Object Detection via Sensitivity-Guided Pruning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that continual test-time object detection improves when domain-sensitive feature channels are pruned rather than adapted, raising ten-round average mAP by up to 2.3 points over the previous best method while cutting…

desk verdict Useful engineering contribution to CTTA-OD, but the paper overclaims its sensitivity-guiding mechanism without a random-pruning control or error bars. read the letter →

arxiv 2506.02462 v1 pith:7IRWU5UH submitted 2025-06-03 cs.CV

Kunyu Wang , Xueyang Fu , Xin Lu , Chengjie Ge , Chengzhi Cao , Wei Zhai , Zheng-Jun Zha This is my paper
classification cs.CV
keywords continualtest-timeadaptationobjectdetectionchannelpruningdomainshiftbatchnormalizationscalingfactorsfeaturealignmentefficiencysensitivityanalysis
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

This paper claims that continual test-time adaptation of object detectors can be made both more accurate and cheaper by pruning away feature channels that are sensitive to domain shift, rather than adapting all channels uniformly. Using the scaling factors of batch-normalization layers as pruning handles, the method assigns each channel a sensitivity weight from how much its image-level and instance-level features deviate from pre-computed source statistics, and applies weighted sparsity to suppress the most sensitive channels. A stochastic reactivation mechanism lets pruned channels return so early pruning mistakes can be corrected. On three benchmark suites (Cityscapes-C, ACDC, UAVDT-C) the method raises ten-round average mAP by 2.3, 2.3, and 0.6 over the prior state of the art while cutting total FLOPs by about 11–13%. If correct, this shows that selectively adapting invariant features is not just faster but also better, offering a new structural route to efficient online adaptation.

What carries the argument

The load-bearing object is the batch-normalization scaling factor $\gamma_i$ used as a per-channel pruning handle, following the network-slimming approach. Sensitivity is quantified by $\omega = \omega_{\mathrm{img}} + \omega_{\mathrm{ins}}$: $\omega_{\mathrm{img}}$ averages channel-wise L1 distance between target feature maps and pre-computed source average features, normalized by channel sum; $\omega_{\mathrm{ins}}$ does the same over RoI-aligned instance features. The weighted sparsity loss $\mathcal{L}_{\mathrm{wreg}} = \sum_i \|\omega_i \cdot \gamma_i\|_1$ pushes sensitive channels' scaling factors toward zero; channels with $\gamma < t$ are excised from the forward/backward graph, and a Bernoulli reactivation with probability $r$ restores some pruned channels to their pre-trained scaling values. This mechanism is what lets the method concentrate adaptation on invariant channels and shrink compute.

What would settle it

Run the method on all three benchmarks with the sensitivity weights replaced by random per-channel weights at the same pruning ratio; if random pruning matches sensitivity-guided pruning in mAP and FLOPs, the sensitivity signal is not doing the work. Alternatively, measure the correlation between each channel's sensitivity score and the per-channel removal effect on cross-domain mAP across multiple backbones and shift types; a weak or negative correlation would falsify the core assumption.

Watch

Extended reading notes

Core claim

The paper's central claim is that in a source-pretrained detector, some feature channels are 'domain-sensitive': they help on the source distribution but systematically hurt on shifted target distributions, so they should be identified and pruned during test-time adaptation. The evidence is a scatter-plot experiment where removing certain channels improves cross-domain mAP while lowering in-domain mAP. The proposed method turns this observation into a training objective: a sensitivity weight per batch-normalization channel, derived from L1 deviation between target features (whole images and RoI-aligned instances) and pre-computed source statistics, multiplies the sparsity regularizer on the BN scaling factors, so high-sensitivity channels are pushed below a pruning threshold and removed from the computation graph. The reported result is that pruning these channels yields higher average mAP over ten rounds of continual adaptation than adapting all channels, while total FLOPs fall by 11.6%, 12.9%, and 10.9% on the three benchmarks.

Load-bearing premise

The load-bearing premise is that channels whose activations deviate most from source statistics under target data are the channels that hurt cross-domain accuracy, so pruning them helps; the paper supports this with a qualitative scatter plot on one corruption benchmark and one detector, not with a quantitative or theoretical proof.

Editorial extensions

If this is right

  • Continual test-time adaptation can be made cheaper without sacrificing accuracy: the same detector, pruned during adaptation, uses fewer FLOPs in both forward and backward passes.
  • Adapting only invariant channels reduces the risk of error accumulation over multiple rounds, since fewer parameters are updated on noisy target signals.
  • The pruning ratio threshold $p$ acts as a budget control, letting practitioners trade accuracy for compute by setting how many channels may be removed.
  • The method updates only batch-normalization scaling factors using pre-computed source statistics, so source images are never needed during adaptation.
  • The gains are measured across autonomous driving and UAV benchmarks, suggesting applicability to resource-constrained perception systems.

Reading between the lines

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

  • The sensitivity score could be computed online from a running estimate of target statistics, avoiding the need for pre-computed source features; the paper does not test this variant.
  • If the sensitivity–harmfulness correlation holds, the same pruning principle might transfer to other self-training or alignment-based CTTA-OD losses, not just the KL-divergence alignment used here.
  • The reported FLOPs savings are structural; actual wall-clock speedup depends on whether the sparsity is exploited by the inference engine, which the paper does not measure.
  • Testing whether the same red-point pattern in the scatter-plot evidence appears for other backbones and shift types would directly strengthen or bound the method's scope.
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

4 major / 6 minor

Summary. The manuscript proposes an efficient continual test-time adaptive object detection method based on channel pruning. A sensitivity weight for each BN channel is derived from image-level and instance-level feature deviations between target and pre-computed source statistics (Eqs. 5-8). The weighted L1 sparsity loss suppresses high-sensitivity channels, with hard pruning when the BN scale falls below a threshold and stochastic reactivation to avoid irreversible loss. Experiments on Cityscapes->Cityscapes-C, UAVDT->UAVDT-C, and Cityscapes->ACDC with Faster R-CNN/ResNet18 report average mAP increases of 2.3, 2.3, and 0.6 over the prior SOTA (WHW) at 11-13% lower FLOPs.

Significance. If the results are robust, the paper makes a useful contribution: it is among the first to treat structural pruning as a way to improve both accuracy and efficiency in CTTA-OD, and the stochastic reactivation is a sensible countermeasure to irreversible pruning. The method is clearly specified and the ablations show consistent gains from the main components. The paper's strengths include a three-benchmark evaluation, a consistent improvement over five baselines, and an explicit FLOPs comparison. However, the evidence for the core mechanism (that feature-deviation sensitivity identifies harmful channels) is qualitative, and the empirical claims lack error bars and a random-pruning control, so the significance hinges on additional validation.

major comments (4)
  1. [Section 4.3, Table 4] The ablation does not isolate sensitivity guidance from pruning itself. The row with L_adp+L_reg (uniform sparsity) achieves 8.8 mAP with 10.2% FLOPs reduction, while adding both sensitivity terms reaches 11.2 mAP with 10.7% FLOPs reduction; however, there is no control that prunes the same number of randomly selected channels at the same FLOPs. Without such a control, the improved mAP could be attributed to pruning plus the adaptation loss rather than to the sensitivity weighting. Moreover, rows are single runs with no error bars. Please add an equal-FLOPs random-pruning baseline and, if possible, a control that uses the same sensitivity scores but prunes low-sensitivity channels (inverted ranking) to demonstrate the direction of the effect.
  2. [Section 4.1, Tables 1-3] All comparisons are single runs; no standard deviation or significance test is reported for the mAP values. The improvement over WHW in Table 3 is only 0.6 mAP (24.2 vs 23.6), and the per-condition differences are within the typical run-to-run variability of stochastic adaptation. In addition, the hyperparameters t=0.05, p=0.1, r=0.01, and lambda=0.05 are fixed without a described validation split; Fig. 5 sweeps each hyperparameter on Cityscapes-C only, which does not establish that the selected values generalize. Please report mean and standard deviation over at least three random seeds and describe the hyperparameter selection protocol.
  3. [Section 4.1 and Section 3.3] The efficiency claim is based on FLOPs counted only for forward and backward passes of the network. The computation of the sensitivity weights (Eqs. 5-8) involves, for every adaptation step, per-image absolute differences on feature maps and per-RoI feature extraction and alignment; these operations are outside the reported Fwd/Bwd FLOPs. Since the stated contribution is efficiency, the overhead should either be included in the FLOPs total or wall-clock time per adaptation round should be reported. Otherwise the reported 10.9-12.9% FLOPs reductions over WHW may overstate the practical efficiency gain.
  4. [Section 3.3 and Figure 2] The load-bearing premise is that channels whose target features deviate most from pre-computed source statistics are the channels that hurt cross-domain performance. The only direct evidence is Fig. 2, a qualitative scatter plot for one detector (ResNet18/Faster R-CNN) on one corruption benchmark (Cityscapes-C), and no quantitative correlation coefficient is reported between the proposed omega and the measured per-channel cross-domain impact. Moreover, the scatter plot is produced by removing single channels from the frozen source model, whereas during adaptation the gamma scales are jointly optimized with the sparsity loss; static removal need not predict the effect of penalizing gamma in the online optimization. Please provide a quantitative correlation analysis on at least two backbones/corruption sets, or an experiment that shows the sensitivity ranking is stable across target domains.
minor comments (6)
  1. [Table 4] The checkmark columns are not aligned with the rows; please reformat so each row unambiguously indicates which components are active.
  2. [Figure 2] Specify how a channel is 'removed' (e.g., setting gamma=0 or deleting the filter) and define the axes more precisely, since the current caption does not explain the red/blue point separation quantitatively.
  3. [Section 3.4, Eq. (9)] The KL divergence uses the source covariance Sigma_s for both Gaussian terms, so it reduces to a Mahalanobis mean-matching term and does not align target variance; please justify this choice or estimate a target covariance.
  4. [Algorithm 1] Clarify the order of hard pruning (line 2) and reactivation (line 9) within one step, and state whether the pruned computation graph is rebuilt after reactivation.
  5. [Section 4.1] Please state explicitly which baseline numbers are re-implemented and whether code/checkpoints will be released.
  6. [Abstract and Section 4.2] The phrase 'the recent SOTA method' should name WHW explicitly in both places for reproducibility and clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: sensitivity weights are online feature-discrepancy statistics, evaluated on external benchmarks.

full rationale

The claimed derivation chain is self-contained. Sensitivity weights (Eqs. 5-8) are computed as channel-wise normalized L1 discrepancies between current target features and pre-computed source statistics; these weights enter a weighted sparsity loss (Eq. 3) and a pruning rule (Algorithm 1, gamma < t). This is an online adaptation signal, not a fitted parameter later renamed as a prediction. The reported gains (2.3/2.3/0.6 mAP, Tables 1-3) are measured on three external benchmarks against independently published baselines, and Table 4 includes a uniform-pruning control. The motivation in Fig. 2 is a qualitative, single-setting observation, and Table 4 lacks a random-pruning control; those are empirical-support weaknesses, not circularity. Hyperparameters (t, p, r, lambda) are selected settings rather than fitted outputs, so selecting them on test data would be a tuning concern, not a definitional reduction. Self-citations (e.g., refs. 14, 15, 39, 40, 45, 46) concern prior UAV/detection work and are not load-bearing for the pruning losses or the evaluation. The Conclusion's stated limitations (BN-only detectors, general invariant features) are explicit scope limits, not circular dependencies. No equation or cited theorem equates the claimed outcome with an input quantity. Therefore no circular step is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method relies on manually chosen hyperparameters (t, p, r, lambda) and several domain assumptions about channel importance, source statistics, and the relationship between feature discrepancy and channel harmfulness. No new physical or conceptual entities are introduced.

free parameters (4)
  • Pruning threshold t = 0.05
    Channels with BN scaling factor below t are pruned. Chosen by hand; no cross-validation described.
  • Pruning ratio threshold p = 0.1
    When the pruned ratio exceeds p, the sparsity loss is disabled. Chosen by hand.
  • Reactivation probability r = 0.01
    Bernoulli probability for reactivating a pruned channel. Chosen by hand.
  • Loss balancing coefficient lambda = 0.05
    Weight for the weighted sparsity loss in the total loss. Chosen by hand.
assumptions (4)
  • domain assumption BN scaling factors are valid indicators of channel importance for pruning
    Adopted from network slimming (Liu et al. 2017); the paper does not re-validate this on the detection backbone.
  • domain assumption Channels with high target-source feature discrepancy are harmful to target-domain performance
    Supported only by a qualitative exploratory experiment on Cityscapes-C (Fig. 2); no quantitative analysis or statistical test.
  • domain assumption Pre-computed source feature statistics are available and accurately represent the source distribution
    Standard in TTA literature; the paper does not describe how these are computed or how sensitive results are to them.
  • standard math ResNet convolutions between two BN layers support the claimed quadratic FLOPs reduction
    The authors discuss this in Section 3.4; it follows from the structure of ResNet but is not formally proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Test-time Adaptive Object Detection via Sensitivity-Guided Pruning." pith.science (2026). https://pith.science/paper/7IRWU5UH

@misc{pith2026250602462,
  author       = {Pith},
  title        = {Pith review of: Efficient Test-time Adaptive Object Detection via Sensitivity-Guided Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7IRWU5UH}},
  note         = {Machine review of arXiv:2506.02462}
}
read the original abstract

Continual test-time adaptive object detection (CTTA-OD) aims to online adapt a source pre-trained detector to ever-changing environments during inference under continuous domain shifts. Most existing CTTA-OD methods prioritize effectiveness while overlooking computational efficiency, which is crucial for resource-constrained scenarios. In this paper, we propose an efficient CTTA-OD method via pruning. Our motivation stems from the observation that not all learned source features are beneficial; certain domain-sensitive feature channels can adversely affect target domain performance. Inspired by this, we introduce a sensitivity-guided channel pruning strategy that quantifies each channel based on its sensitivity to domain discrepancies at both image and instance levels. We apply weighted sparsity regularization to selectively suppress and prune these sensitive channels, focusing adaptation efforts on invariant ones. Additionally, we introduce a stochastic channel reactivation mechanism to restore pruned channels, enabling recovery of potentially useful features and mitigating the risks of early pruning. Extensive experiments on three benchmarks show that our method achieves superior adaptation performance while reducing computational overhead by 12% in FLOPs compared to the recent SOTA method.

Figures

Figures reproduced from arXiv: 2506.02462 by the authors.

Figure 1
Figure 1. We consider continual test-time adaptive object detec [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. We perform continual online adaptation on Cityscapes [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Overview of our method. Motivated by the observation that certain feature channels learned from the source domain negatively [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Ablation analysis of three hyper-parameters: (a) Pruning threshold [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visualization analysis of the effectiveness of pruning sensitive channels. (a) mAP change curves during test-time adaptation for [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 36 canonical work pages

  1. [1]

    Energy effi- cient uav-enabled mobile edge computing for iot devices: A review.IEEE Access, 9:127779–127798, 2021

    Muhammad Abrar, Ushna Ajmal, Ziyad M Almohaimeed, Xiang Gui, Rizwan Akram, and Roha Masroor. Energy effi- cient uav-enabled mobile edge computing for iot devices: A review.IEEE Access, 9:127779–127798, 2021. 1

    work page 2021

  2. [2]

    Continual Test-Time Adaptation for Object Detection with Adaptive Monitoring and Randomized Restoration

    Shilei Cao, Yan Liu, Juepeng Zheng, Weijia Li, Runmin Dong, and Haohuan Fu. Exploring test-time adaptation for object detection in continually changing environments.arXiv preprint arXiv:2406.16439, 2024. 3

    work page Pith review arXiv 2024

  3. [3]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. InEuropean confer- ence on computer vision, pages 213–229. Springer, 2020. 8

    2020

  4. [4]

    Stfar: Im- proving object detection robustness at test-time by self- training with feature alignment regularization.arXiv preprint arXiv:2303.17937, 2023

    Yijin Chen, Xun Xu, Yongyi Su, and Kui Jia. Stfar: Im- proving object detection robustness at test-time by self- training with feature alignment regularization.arXiv preprint arXiv:2303.17937, 2023. 1, 3, 4, 5, 6, 7

    arXiv 2023

  5. [5]

    Exploiting low-confidence pseudo-labels for source-free object detec- tion

    Zhihong Chen, Zilei Wang, and Yixin Zhang. Exploiting low-confidence pseudo-labels for source-free object detec- tion. InProceedings of the 31st ACM International Confer- ence on Multimedia, pages 5370–5379, 2023. 2

    work page 2023

  6. [6]

    A survey on deep neural network pruning: Taxonomy, compar- ison, analysis, and recommendations.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    Hongrong Cheng, Miao Zhang, and Javen Qinfeng Shi. A survey on deep neural network pruning: Taxonomy, compar- ison, analysis, and recommendations.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3

    work page 2024

  7. [7]

    Adversarial alignment for source free object detection

    Qiaosong Chu, Shuyan Li, Guangyi Chen, Kai Li, and Xiu Li. Adversarial alignment for source free object detection. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 452–460, 2023. 2

    work page 2023

  8. [8]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016. 6

    work page 2016

Show all 47 references
  1. [9]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 6

    2009

  2. [10]

    Balanced teacher for source-free object detection.IEEE Transactions on Circuits and Systems for Video Technology, 2024

    Jinhong Deng, Wen Li, and Lixin Duan. Balanced teacher for source-free object detection.IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2

    2024

  3. [11]

    Ob- ject detection using yolo: Challenges, architectural succes- sors, datasets and applications.multimedia Tools and Appli- cations, 82(6):9243–9275, 2023

    Tausif Diwan, G Anirudh, and Jitendra V Tembhurne. Ob- ject detection using yolo: Challenges, architectural succes- sors, datasets and applications.multimedia Tools and Appli- cations, 82(6):9243–9275, 2023. 8

    2023

  4. [12]

    The unmanned aerial vehicle benchmark: Object detection and tracking

    Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian. The unmanned aerial vehicle benchmark: Object detection and tracking. InProceedings of the European con- ference on computer vision (ECCV), pages 370–386, 2018. 6

    2018

  5. [13]

    Dynamic channel pruning: Feature boosting and suppression.arXiv preprint arXiv:1810.05331,

    Xitong Gao, Yiren Zhao, Łukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression.arXiv preprint arXiv:1810.05331,

    arXiv

  6. [14]

    Learning dual convolutional dictionaries for image de-raining

    Chengjie Ge, Xueyang Fu, and Zheng-Jun Zha. Learning dual convolutional dictionaries for image de-raining. InACM MM, pages 6636–6644, 2022. 1

    2022

  7. [15]

    Neuromorphic event signal-driven network for video de-raining

    Chengjie Ge, Xueyang Fu, Peng He, Kunyu Wang, Chengzhi Cao, and Zheng-Jun Zha. Neuromorphic event signal-driven network for video de-raining. InAAAI, pages 1878–1886,

  8. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 6

    2016

  9. [17]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. InProceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. 4

    2017

  10. [18]

    Structured pruning for deep con- volutional neural networks: A survey.IEEE transactions on pattern analysis and machine intelligence, 2023

    Yang He and Lingao Xiao. Structured pruning for deep con- volutional neural networks: A survey.IEEE transactions on pattern analysis and machine intelligence, 2023. 3

    2023

  11. [19]

    Soft filter pruning for accelerating deep convolutional neural networks.arXiv preprint arXiv:1808.06866, 2018

    Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks.arXiv preprint arXiv:1808.06866, 2018. 3

    2018 arXiv

  12. [20]

    Benchmarking neu- ral network robustness to common corruptions and perturba- tions.arXiv preprint arXiv:1903.12261, 2019

    Dan Hendrycks and Thomas Dietterich. Benchmarking neu- ral network robustness to common corruptions and perturba- tions.arXiv preprint arXiv:1903.12261, 2019. 6

    1903 arXiv

  13. [21]

    What’s the backward- forward flop ratio for neural networks?, 2021

    Marius Hobbhahn and Jaime Sevilla. What’s the backward- forward flop ratio for neural networks?, 2021. Accessed: 2024-10-26. 6

    2021

  14. [22]

    Channel gating neural networks.Ad- vances in Neural Information Processing Systems, 32, 2019

    Weizhe Hua, Yuan Zhou, Christopher M De Sa, Zhiru Zhang, and G Edward Suh. Channel gating neural networks.Ad- vances in Neural Information Processing Systems, 32, 2019. 3

    2019

  15. [23]

    Dynamic retraining-updating mean teacher for source-free object de- tection.arXiv preprint arXiv:2407.16497, 2024

    Trinh Le Ba Khanh, Huy-Hung Nguyen, Long Hoang Pham, Duong Nguyen-Ngoc Tran, and Jae Wook Jeon. Dynamic retraining-updating mean teacher for source-free object de- tection.arXiv preprint arXiv:2407.16497, 2024. 2

    2024 arXiv

  16. [24]

    Dynamic slimmable network

    Changlin Li, Guangrun Wang, Bing Wang, Xiaodan Liang, Zhihui Li, and Xiaojun Chang. Dynamic slimmable network. InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 8607–8617, 2021. 3

    2021

  17. [25]

    A free lunch for unsuper- vised domain adaptive object detection without source data

    Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsuper- vised domain adaptive object detection without source data. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 8474–8481, 2021. 2

    2021

  18. [26]

    Runtime neural pruning.Advances in neural information processing systems, 30, 2017

    Ji Lin, Yongming Rao, Jiwen Lu, and Jie Zhou. Runtime neural pruning.Advances in neural information processing systems, 30, 2017. 3

    2017

  19. [27]

    Periodically exchange teacher-student for source-free object detection

    Qipeng Liu, Luojun Lin, Zhifeng Shen, and Zhifeng Yang. Periodically exchange teacher-student for source-free object detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 6414–6424, 2023. 2

    2023

  20. [28]

    Edge computing for autonomous driving: Opportunities and challenges.Proceedings of the IEEE, 107 (8):1697–1716, 2019

    Shaoshan Liu, Liangkai Liu, Jie Tang, Bo Yu, Yifan Wang, and Weisong Shi. Edge computing for autonomous driving: Opportunities and challenges.Proceedings of the IEEE, 107 (8):1697–1716, 2019. 1

    2019

  21. [29]

    Learning efficient convolutional networks through network slimming

    Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. InPro- ceedings of the IEEE international conference on computer vision, pages 2736–2744, 2017. 3, 4

    2017

  22. [30]

    The norm must go on: Dynamic unsuper- vised domain adaptation by normalization

    M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsuper- vised domain adaptation by normalization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14765–14775, 2022. 3, 6, 7

    2022

  23. [31]

    Act- mad: Activation matching to align distributions for test- time-training

    Muhammad Jehanzeb Mirza, Pol Jan ´e Soneira, Wei Lin, Ma- teusz Kozinski, Horst Possegger, and Horst Bischof. Act- mad: Activation matching to align distributions for test- time-training. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page...

    2023

  24. [32]

    Faster r-cnn: Towards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.IEEE transactions on pattern analysis and machine intelligence, 39(6):1137–1149, 2016. 6

    2016

  25. [33]

    Fully test-time adaptation for object detection

    Xiaoqian Ruan and Wei Tang. Fully test-time adaptation for object detection. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 1038–1047, 2024. 3

    2024

  26. [34]

    Acdc: The adverse conditions dataset with correspondences for se- mantic driving scene understanding

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for se- mantic driving scene understanding. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765–10775, 2021. 6

    2021

  27. [35]

    Test: Test-time self-training under distribution shift

    Samarth Sinha, Peter Gehler, Francesco Locatello, and Bernt Schiele. Test: Test-time self-training under distribution shift. InProceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision, pages 2759–2769, 2023. 1, 3, 6, 7

    2023

  28. [36]

    A comprehensive review of recent research trends on unmanned aerial vehicles (uavs).Systems, 11(8):400,

    Khaled Telli, Okba Kraa, Yassine Himeur, Abdelmalik Oua- mane, Mohamed Boumehraz, Shadi Atalla, and Wathiq Mansoor. A comprehensive review of recent research trends on unmanned aerial vehicles (uavs).Systems, 11(8):400,

  29. [37]

    Object recognition and detection with deep learning for autonomous driving applications.Simulation, 93(9):759–769, 2017

    Ays ¸eg¨ul Uc ¸ar, Yakup Demir, and C¨uneyt G ¨uzelis ¸. Object recognition and detection with deep learning for autonomous driving applications.Simulation, 93(9):759–769, 2017. 1

    2017

  30. [38]

    Instance rela- tion graph guided source-free domain adaptive object detec- tion

    Vibashan VS, Poojan Oza, and Vishal M Patel. Instance rela- tion graph guided source-free domain adaptive object detec- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 3520–3530,

  31. [39]

    Generalized uav object detec- tion via frequency domain disentanglement

    Kunyu Wang, Xueyang Fu, Yukun Huang, Chengzhi Cao, Gege Shi, and Zheng-Jun Zha. Generalized uav object detec- tion via frequency domain disentanglement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1064–1073, 2023. 1

    2023

  32. [40]

    Towards generalized uav object detection: A novel perspective from frequency domain disentanglement

    Kunyu Wang, Xueyang Fu, Chengjie Ge, Chengzhi Cao, and Zheng-Jun Zha. Towards generalized uav object detection: A novel perspective from frequency domain disentanglement. International Journal of Computer Vision, pages 1–29, 2024. 1

    2024

  33. [41]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201–7211, 2022. 1, 6

    2022

  34. [42]

    Jayeon Yoo, Dongkwan Lee, Inseop Chung, Donghyun Kim, and Nojun Kwak. What how and when should object detec- tors update in continually changing test domains? InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23354–23363, 2024. 3, 4, 5, 6, 7

    2024

  35. [43]

    Enhancing source- free domain adaptive object detection with low-confidence pseudo label distillation.arXiv preprint arXiv:2407.13524,

    Ilhoon Yoon, Hyeongjun Kwon, Jin Kim, Junyoung Park, Hyunsung Jang, and Kwanghoon Sohn. Enhancing source- free domain adaptive object detection with low-confidence pseudo label distillation.arXiv preprint arXiv:2407.13524,

    arXiv

  36. [44]

    Gate decorator: Global filter pruning method for ac- celerating deep convolutional neural networks.Advances in neural information processing systems, 32, 2019

    Zhonghui You, Kun Yan, Jinmian Ye, Meng Ma, and Ping Wang. Gate decorator: Global filter pruning method for ac- celerating deep convolutional neural networks.Advances in neural information processing systems, 32, 2019. 3

    2019

  37. [45]

    Uni-navid: A video-based vision- language-action model for unifying embodied navigation tasks.arXiv preprint arXiv:2412.06224, 2024

    Jiazhao Zhang, Kunyu Wang, Shaoan Wang, Minghan Li, Haoran Liu, Songlin Wei, Zhongyuan Wang, Zhizheng Zhang, and He Wang. Uni-navid: A video-based vision- language-action model for unifying embodied navigation tasks.arXiv preprint arXiv:2412.06224, 2024. 1

    2024 arXiv

  38. [46]

    Navid: Video-based vlm plans the next step for vision-and-language navigation.arXiv preprint arXiv:2402.15852, 2024

    Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and He Wang. Navid: Video-based vlm plans the next step for vision-and-language navigation.arXiv preprint arXiv:2402.15852, 2024. 1

    2024 arXiv

  39. [47]

    Multi-source-free domain adaptive object detection.International Journal of Computer Vision, pages 1–33, 2024

    Sicheng Zhao, Huizai Yao, Chuang Lin, Yue Gao, and Guiguang Ding. Multi-source-free domain adaptive object detection.International Journal of Computer Vision, pages 1–33, 2024. 2

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.