Pith. sign in

REVIEW 4 major objections 7 minor 85 references

Mitigating Context Bias in Domain Adaptation for Object Detection using Mask Pooling

T0 review · 4 major / 7 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 The paper claims that standard pooling layers are a hidden source of context bias in object detection, and that replacing them with mask-conditioned pooling removes spurious foreground-background associations and makes detectors robust to…

desk verdict A useful benchmark and a simple pooling trick, but the causal claim is undermined by a confound: mask-pooled models get oracle FG masks at inference and the baselines do not. read the letter →

arxiv 2505.18446 v1 pith:6EOCYDSD submitted 2025-05-24 cs.CV cs.AI

Hojun Son , Asma Almutairi , Arpan Kusari This is my paper
classification cs.CVcs.AI
keywords contextbiasdomainadaptationobjectdetectionmaskpoolingcausalinterventionforegroundmasksrandombackgroundbenchmarkspuriouscorrelation
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 claims that the pooling layers inside convolutional detectors are a hidden source of context bias: by aggregating features without distinguishing foreground from background, max pooling can encode spurious FG-BG associations that break when the model meets a new domain. To test this, the authors intervene on the pooling operation directly, replacing it with Mask Pooling, which uses ground-truth foreground masks to average separately over foreground and background pixels and then picks the majority region's average. Across Cityscapes and Virtual KITTI, including synthetic images where backgrounds are replaced by random samples from BG-20K, the masked models detect objects more robustly and show lower performance variation under background perturbation. If right, this reframes DAOD as a problem at the feature-extraction layer rather than only a training-strategy problem.

What carries the argument

Mask Pooling (Eq. 1): a kernel-wise pooling rule that, given a binary foreground mask F, outputs the average of the foreground pixel values when the kernel contains at least as many foreground as background pixels, and the average of the background pixel values otherwise—i.e., a mask-conditioned majority-average pooling. It carries the causal intervention: by making the pooling outcome depend explicitly on the FG/BG split, it blocks the 'F → A ← B' v-structure that the paper identifies as the source of context bias. This operation is inserted in place of or alongside existing pooling layers in the detector backbones.

What would settle it

Train the same detector with standard max pooling, mask pooling, and plain average pooling (no mask) on Cityscapes, then evaluate on random-background synthetic images from BG-20K. If plain average pooling performs as well as mask pooling, the robustness gain is not caused by separating foreground from background, and the proposed causal mechanism is not supported.

Watch

Extended reading notes

Core claim

Standard pooling operations, especially max pooling, indiscriminately select or average activations across both foreground and background pixels, creating an implicit association node 'A' between FG and BG that drives predictions and breaks under domain shift. Mask Pooling removes this association by conditioning the pooling outcome on a binary foreground mask: within each 3x3 kernel, it averages the foreground pixels when they are the majority and the background pixels otherwise. The paper reports that inserting this layer into ResNet-50 and EfficientNet-B0 based detectors improves mAP50 on Cityscapes and Virtual KITTI and, more strongly, on synthetic benchmarks with random backgrounds, while also improving hierarchical F1 robustness during background activation perturbations. The authors argue this is a causal intervention that forces the detector to rely on foreground features rather than context.

Load-bearing premise

The method requires accurate foreground masks for every training and test image, and the paper's own ablation shows that boundary errors of about 20% cause large performance drops; without reliable masks, the reported robustness gains do not transfer to real deployment.

Editorial extensions

If this is right

  • Detectors using Mask Pooling maintain higher mAP50 than their baseline and often outperform ALDI++, a target-domain-aware method, when backgrounds are randomized, despite receiving no target-domain information.
  • EfficientNet-based models, which lack early max pooling, show the largest hierarchical F1 gains from the intervention (76 of 77 evaluation pairs), suggesting the benefit scales with how much pooling the architecture relies on.
  • The random-background benchmark (FG objects composited onto BG-20K images) is proposed as an 'ultimate test' for DAOD: any detector that relies on spurious BG cues should fail it, while FG-conditioned models should pass.
  • Replacing pooling with mask pooling requires no change to the loss function or training schedule, only the extra mask input at the pooling location, so it combines with existing DAOD training strategies.

Reading between the lines

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

  • The paper's own ablation shows that a 20% boundary error in the masks drops mAP from about 73.6 to 53.8 on Cityscapes; a natural extension is to couple Mask Pooling with a mask predictor (e.g., a pretrained segmenter) and measure how much of the robustness gain survives imperfect masks at inference.
  • If the causal story is right, the same intervention should help other tasks where foreground/background separation is well-defined, such as instance segmentation or visual inspection, not just DAOD.
  • The random-background benchmark could be turned into a standardized stress test: report mAP50 and its variance across many random BG samples, which directly exposes whether a detector is modeling context rather than objects.
  • The authors' limitation note frames a chicken-and-egg loop: masks can be generated from learned features, or features can be separated by masks. A testable extension is to train Mask Pooling with masks predicted by an off-the-shelf segmenter and fine-tune jointly, checking if the intervention remains beneficial when masks are imperfect.
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 / 7 minor

Summary. The paper argues that context bias in domain-adaptive object detection arises from standard pooling operations that indiscriminately mix foreground and background features. It proposes Mask Pooling (Eq. 1), which uses ground-truth foreground masks to compute separate average-pooled values for foreground and background regions, selecting one based on a majority vote. The authors claim this implements a causal intervention on an implicit association variable A, thereby making detectors robust to background changes and domain shift. They report consistent mAP50 improvements for mask-pooled ResNet-50 and EfficientNet-B0 over their unmodified baselines and the ALDI++ DAOD method across Cityscapes, Cityscapes foggy/rainy, KITTI, Virtual KITTI, and synthetic random-background benchmarks, and they introduce a new benchmark with randomly sampled backgrounds.

Significance. If established, the claim that pooling layers are a causal source of context bias would offer a simple architectural principle for improving object-detection robustness, and the proposed Mask Pooling is easy to integrate into existing backbones. The paper also contributes an extensive multi-dataset evaluation and a new synthetic benchmark. However, the significance is currently undercut by a central confound: the mask-pooled models receive oracle foreground masks at inference while the baselines do not, so the experimental design does not isolate the pooling operation as the causal agent. The causal narrative in Section 2 is asserted rather than derived, and Section 5.1 concedes that the method depends on masks, which conflicts with the claimed unsupervised DAOD setting. The reported gains may reflect the effect of mask information rather than a property of the pooling mechanism.

major comments (4)
  1. [Section 3.1, Eq. (1); Tables 2, 3, 6, 7] The experimental comparison does not isolate the pooling operation as the cause of the reported gains. Mask-pooled models (ResM, EffM) receive ground-truth foreground masks at both training and inference, while the baselines (Res, Eff, ALDI++) receive no such signal. Thus, the mAP50 improvements could equally be explained by the mask information alone, which at the early feature level is a strong cue about object locations. The same-domain result on Cityscapes validation (Table 2: ResM 73.572 vs Res 65.386) is particularly telling because there is no domain shift to remove. To support the central causal claim, the paper needs a control condition that supplies the same mask information to the baseline pooling through a different mechanism (e.g., mask-weighted max pooling or a separate mask-conditioned feature branch). Without such a control, Eq. (1) changes both the pooling operation and the information available to the network, so the assertion that pooling operations introduce causal artifacts is not established.
  2. [Section 2] The causal derivation is a factorization, not a derivation of do(A). The paper writes P(Y,F|do(A)) = P(Y|F, do(A))P(F) and then approximates it by P(Y,F)≈P(Y|F)P(F), but the variable A is never formally defined and the mapping from Eq. (1) to a do-operation is not shown. Simplifying P(Y|F, do(A)) to P(Y|F) assumes away the very association the paper claims to remove, making the causal narrative partly self-referential: the method defines the intervention, and the benchmark is constructed with the same masks. Please either formalize the structural causal model and explicitly derive what do(A) means in terms of the network computation, or relax the causal language to a mechanistic hypothesis and present the masks as an informative auxiliary input.
  3. [Section 5.1; Sections 3.4.2, 3.4.3] The acknowledged reliance on foreground masks at both training and inference conflicts with the paper's framing as a source-only DAOD solution. Because target-domain masks are unavailable in the standard unsupervised DAOD setting, the reported robustness gains do not transfer to the method's intended use case; Section 4.6 shows that even 10% boundary error causes large performance drops. Moreover, the synthetic random-background benchmark is constructed with the same ground-truth masks, so it is tailored to mask-informed models and does not independently validate the pooling mechanism. This limitation is not merely a deployment issue; it is a confound in the headline comparisons because the baselines never receive the privileged mask signal.
  4. [Tables 4 and 5] Hierarchical F1 scores for EffM on VKC Truck are reported as 1.136580, which exceeds the maximum possible F1 score of 1. This suggests an error in the computation or the table formatting. The claim that EffM improved in 76 out of 77 evaluation pairs depends on these numbers, so they should be verified and corrected. Please also clarify the exact definition of the hierarchical F1 score used and how it is aggregated across classes and datasets.
minor comments (7)
  1. [Abstract / Introduction] The introduction mentions 'source-only domain' as if the method requires no target information, but the method uses target-domain foreground masks during inference. Please clarify the intended setting and avoid this contradiction.
  2. [Section 3.1] Eq. (1) describes a 3x3 kernel region, but the stride, padding, and downsampling behavior of the mask pooling layer are not specified. Please state these details so the layer can be reproduced.
  3. [Section 3.2] For ResNet-50, the text says max pooling is 'replaced with a mask pooling operation inserted after the first block'; please clarify whether the original max pooling is removed entirely or the mask pooling is added in parallel, since this affects the information flow and the parameter/flop comparison.
  4. [Section 4.3] The sentence 'Even EffM outperformed ALDI++ by about 2% with lowest computational costs among other models (see Table 2)' is not supported by Table 2 as printed: EffM is lower than ALDI++ on CV, CFV, CRV, and KST except for KST (50.027 vs 47.965). Please rephrase to specify the dataset and provide the exact comparison.
  5. [Section 4.6] The ablation text reports a performance of 73.578 on CV, while Table 2 reports 73.572 for ResM. Please verify and use a consistent value.
  6. [Table 9] The note that EffM's lowest mAP50 (16.25) is 'disregarded due to lower Max mAP50' is not self-explanatory. Please explain the exclusion criterion so that the robustness comparison is not perceived as cherry-picking.
  7. [References] Several references are malformed or incomplete; for example, reference [1] appears to have broken author formatting. A full proofread of the bibliography and surrounding text is needed.

Circularity Check

2 steps flagged · score 4.0 of 10

Causal claim is partly self-definitional (intervention defined as mask use; benchmark built from the same masks) and the v-structure is imported from the authors' own prior work; numerical gains remain empirical, so score is moderate.

  1. self definitional [Section 2 (Problem definition) and Section 3.1 (Eq. 1); benchmark in Section 3.4.2]
    "Case (3) introduces causal intervention to analyze context bias by intervening on 'A'. ... The use of masks enforces determinism in the variable 'A' by explicitly separating FG from BG, thereby constraining the influence of spurious contextual features. ... The designed pooling separates each region using ground truth masks into FG and BG during training and inference. ... The synthetic datasets consists of randomly chosen BG images compounded with FG images using FG masks."

    The paper defines the causal intervention do(A) as the act of using FG masks, then defines Mask Pooling (Eq. 1) as a pooling operation that consumes those masks. The random-BG benchmark is constructed with the very same FG masks that are given to Mask Pooling at inference. Comparing Res/Eff (no mask) with ResM/EffM (mask) therefore changes both the pooling operation and the availability of an oracle foreground signal simultaneously. Any mAP gain could be produced by the mask alone, so the conclusion that 'pooling operations can introduce causal artifacts' is assumed in the definition of the intervention rather than empirically isolated. The causal attribution is self-referential, even though the mAP numbers themselves are not fitted.

  2. self citation load bearing [Section 2, Problem definition, citation [62]]
    "Case (2) reflects findings from a quantification study in object detection [62], modeled as a v-structure: 'F' → 'A' ← 'B' ... Besides, the study in [62] demonstrated that FG-to-FG features across domains are more similar than BG-related combinations, enabling robustness even target domains are unseen during training."

    The causal v-structure (F → A ← B and F → Y ← A) and the claim that FG-to-FG features transfer across domains are load-bearing premises for the paper's causal interpretation of Mask Pooling. Both are attributed to [62], an arXiv preprint by the same authors (Son and Kusari), and are not re-derived or independently verified in this paper. The conclusion that Mask Pooling 'remove[s] the causal effects of FG-BG associations' inherits these self-cited premises. The numerical comparisons do not depend on [62], but the causal framing does.

full rationale

The paper does not fit any free parameters to the reported mAP/F1 scores, and the empirical comparison includes external baselines (ALDI++) and external datasets (Cityscapes, KITTI, Virtual KITTI, BG-20K). In that sense the core quantitative results are not derived from the method's own assumptions. However, the load-bearing causal narrative is partly self-definitional: Section 2 explicitly equates the intervention do(A) with using FG masks, and Section 3.1 defines Mask Pooling as pooling with those masks. The random-background benchmark is then built from the same FG masks (Section 3.4.2), so the method receives at inference the exact ground-truth object boundaries used to construct the test images. The Res-vs-ResM and Eff-vs-EffM comparisons therefore vary both the pooling operation and the oracle mask signal, preventing the experiments from isolating 'pooling' as the causal source of context bias. Additionally, the v-structure and FG-similarity premises are imported from the authors' own prior work [62] rather than established here. These issues make the causal claim partially circular, but they do not reduce the empirical mAP improvements to a tautology, hence a moderate score of 4 rather than 6-10.

Assumptions & free parameters 1 free parameters · 5 assumptions · 1 invented entities

The method rests on the availability of foreground masks, an assumed causal graph with independent FG and BG, and an unproven identification of mask pooling with do(A). The only hand-chosen design element is the majority criterion in the pooling rule; no scalar parameters are fitted to the reported mAP or F1 values.

free parameters (1)
  • Mask pooling majority criterion (n_F >= n_B) = none, hand-chosen rule
    Equation (1) outputs the foreground average when foreground pixels outnumber background pixels, otherwise the background average. This arbitrary threshold is not derived from theory and directly shapes the claimed intervention.
assumptions (5)
  • standard math Bayes factorization P(Y,A,F,B) = P(Y|F,A) P(A|F,B) P(F) P(B)
    Section 2 defines the joint distribution for the causal graph; the factorization itself is standard, but it is used to support a causal claim that is not derived.
  • domain assumption Foreground and background features are independent causes, P(F)P(B)
    Section 1 states the hypothesis that FG and BG features 'being inherently independent' form spurious associations during training. This independence is assumed, not tested.
  • domain assumption Ground-truth foreground masks are available for all images at training and inference
    All evaluations of ResM and EffM use masks from Cityscapes, KITTI Semantic, Virtual KITTI, or the synthetic benchmark; Section 5.1 acknowledges this as a significant limitation.
  • ad hoc to paper Mask pooling in Eq. (1) realizes the causal intervention do(A)
    Section 2 transitions from P(Y,F|do(A)) to the proposed pooling rule without proving that average-pooling the majority region corresponds to an intervention on the latent association variable A.
  • standard math Anti-causal learning guarantees that maximizing likelihood under the intervention focuses on FG features
    Section 2 cites [36,26] for the anti-causal claim, but the cited result is not stated or verified in the text, and it is not formally connected to mask pooling.
invented entities (1)
  • Latent association variable A
    purpose: Models the unobserved spurious association between foreground F and background B that influences predictions Y in Section 2.
    A is introduced as an unobservable node in the SCM. The paper provides no falsifiable prediction for A outside the mask-pooling intervention, so it functions as a conceptual placeholder for the bias the method claims to remove.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Context Bias in Domain Adaptation for Object Detection using Mask Pooling." pith.science (2026). https://pith.science/paper/6EOCYDSD

@misc{pith2026250518446,
  author       = {Pith},
  title        = {Pith review of: Mitigating Context Bias in Domain Adaptation for Object Detection using Mask Pooling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6EOCYDSD}},
  note         = {Machine review of arXiv:2505.18446}
}
read the original abstract

Context bias refers to the association between the foreground objects and background during the object detection training process. Various methods have been proposed to minimize the context bias when applying the trained model to an unseen domain, known as domain adaptation for object detection (DAOD). But a principled approach to understand why the context bias occurs and how to remove it has been missing. In this work, we provide a causal view of the context bias, pointing towards the pooling operation in the convolution network architecture as the possible source of this bias. We present an alternative, Mask Pooling, which uses an additional input of foreground masks, to separate the pooling process in the respective foreground and background regions and show that this process leads the trained model to detect objects in a more robust manner under different domains. We also provide a benchmark designed to create an ultimate test for DAOD, using foregrounds in the presence of absolute random backgrounds, to analyze the robustness of the intended trained models. Through these experiments, we hope to provide a principled approach for minimizing context bias under domain shift.

Figures

Figures reproduced from arXiv: 2505.18446 by the authors.

Figure 1
Figure 1. Left panel: (1) shows the ideal case where the label “Y" is dependent only on the FG (“F"); (2) presents the actual case where the FG (“F") and BG (“B") gets associated in “A" and influences the prediction; (3) presents our proposed model where the mask pooling removes the association by making the pooling operation separately in FG and BG. Right panel: It shows an example activation map with max pooling and mask po… view at source ↗
Figure 2
Figure 2. Radar chart of mAP50 on Cityscapes and Virtual KITTI. It is normalized by 100 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 52 canonical work pages

  1. [1]

    Augmented reality meets computer vision: Efficient data generation for urban driving scenes.International Journal of Computer Vision (IJCV), 2018

    AlhaijaandHassan, MustikovelaandSiva, MeschederandLars, GeigerandAndreas, and Rotherand- Carsten. Augmented reality meets computer vision: Efficient data generation for urban driving scenes.International Journal of Computer Vision (IJCV), 2018

    2018

  2. [2]

    Contrastive mean teacher for domain adaptive object detectors

    Shengcao Cao, Dhiraj Joshi, Liang-Yan Gui, and Yu-Xiong Wang. Contrastive mean teacher for domain adaptive object detectors. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23839–23848, 2023

    2023

  3. [3]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021

    2021

  4. [4]

    Learning efficient object detection models with knowledge distillation.Advances in neural information processing systems, 30, 2017

    Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation.Advances in neural information processing systems, 30, 2017

    2017

  5. [5]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016

    2016

  6. [6]

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

    Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3339–3348, 2018

    2018

  7. [7]

    Scale- aware domain adaptive faster r-cnn.International Journal of Computer Vision, 129(7):2223– 2243, 2021

    Yuhua Chen, Haoran Wang, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Scale- aware domain adaptive faster r-cnn.International Journal of Computer Vision, 129(7):2223– 2243, 2021

    work page 2021

  8. [8]

    C-cam: Causal cam for weakly supervised semantic segmentation on medical image

    Zhang Chen, Zhiqiang Tian, Jihua Zhu, Ce Li, and Shaoyi Du. C-cam: Causal cam for weakly supervised semantic segmentation on medical image. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11676–11685, 2022

    work page 2022

Show all 85 references
  1. [9]

    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. InProceedings of the IEEE conference on computer vision and pattern recog...

    2016

  2. [10]

    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

    2009

  3. [11]

    Unbiased mean teacher for cross-domain object detection

    Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4091–4101, 2021

    2021

  4. [12]

    Improved regularization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017

    Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017

    2017 arXiv

  5. [13]

    Virtual worlds as proxy for multi-object tracking analysis

    Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 4340–4349, 2016

    2016

  6. [14]

    Unsupervised domain adaptation by backpropagation

    Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. InInternational conference on machine learning, pages 1180–1189. PMLR, 2015

    2015

  7. [15]

    Domain-adversarial training of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1–35, 2016

    2016

  8. [16]

    Lip: Local importance-based pooling

    Ziteng Gao, Limin Wang, and Gangshan Wu. Lip: Local importance-based pooling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3355–3364, 2019. 10

    2019

  9. [17]

    Siou loss: More powerful learning for bounding box regression.arXiv preprint arXiv:2205.12740, 2022

    Zhora Gevorgyan. Siou loss: More powerful learning for bounding box regression.arXiv preprint arXiv:2205.12740, 2022

    2022 arXiv

  10. [18]

    Detectron

    Ross Girshick, Ilija Radosavovic, Georgia Gkioxari, Piotr Dollár, and Kaiming He. Detectron. https://github.com/facebookresearch/detectron, 2018

    2018

  11. [19]

    Uncertainty-aware unsupervised domain adaptation in object detection.IEEE Transactions on Multimedia, 24:2502– 2514, 2021

    Dayan Guan, Jiaxing Huang, Aoran Xiao, Shijian Lu, and Yanpeng Cao. Uncertainty-aware unsupervised domain adaptation in object detection.IEEE Transactions on Multimedia, 24:2502– 2514, 2021

    2021

  12. [20]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. InProceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017

    2017

  13. [21]

    Deep residual learning for image recognition

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

    2016

  14. [22]

    Object detection using sim2real domain randomization for robotic applications.IEEE Transactions on Robotics, 39(2):1225–1243, 2022

    Dániel Horváth, Gábor Erd ˝os, Zoltán Istenes, Tomáš Horváth, and Sándor Földi. Object detection using sim2real domain randomization for robotic applications.IEEE Transactions on Robotics, 39(2):1225–1243, 2022

    2022

  15. [23]

    Mic: Masked image consistency for context-enhanced domain adaptation

    Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11721–11732, 2023

    2023

  16. [24]

    Depth-attentional features for single-image rain removal

    Xiaowei Hu, Chi-Wing Fu, Lei Zhu, and Pheng-Ann Heng. Depth-attentional features for single-image rain removal. InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 8022–8031, 2019

    2019

  17. [25]

    Causal intervention for object detection

    Weiqing Huang, Miao Jiang, Min Li, Bo Meng, Junxing Ren, Shixian Zhao, Ruwen Bai, and Yang Yang. Causal intervention for object detection. In2021 IEEE 33rd International Conference on Tools with Artificial Intelligence (ICTAI), pages 770–774. IEEE, 2021

    2021

  18. [26]

    Semi-supervised interpolation in an anticausal learning scenario.The Journal of Machine Learning Research, 16(1):1923–1948, 2015

    Dominik Janzing and Bernhard Schölkopf. Semi-supervised interpolation in an anticausal learning scenario.The Journal of Machine Learning Research, 16(1):1923–1948, 2015

    1923

  19. [27]

    Decoupled visual causality for robust detec- tion

    Ping Jiang, Xiaoheng Deng, and Shichao Zhang. Decoupled visual causality for robust detec- tion. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

    2023

  20. [28]

    Overview of the accelerated platform for robotics and artificial intelligence nvidia isaac

    Tatiana D Kainova. Overview of the accelerated platform for robotics and artificial intelligence nvidia isaac. In2023 seminar on information computing and processing (ICP), pages 89–93. IEEE, 2023

    2023

  21. [29]

    Algorith- mic recourse under imperfect causal knowledge: a probabilistic approach.Advances in neural information processing systems, 33:265–277, 2020

    Amir-Hossein Karimi, Julius V on Kügelgen, Bernhard Schölkopf, and Isabel Valera. Algorith- mic recourse under imperfect causal knowledge: a probabilistic approach.Advances in neural information processing systems, 33:265–277, 2020

    2020

  22. [30]

    Align and distill: Unifying and improving domain adaptive object detection.Transactions on Machine Learning Research, 2025

    Justin Kay, Timm Haucke, Suzanne Stathatos, Siqi Deng, Erik Young, Pietro Perona, Sara Beery, and Grant Van Horn. Align and distill: Unifying and improving domain adaptive object detection.Transactions on Machine Learning Research, 2025. Featured Certification

    2025

  23. [31]

    The caltech fish counting dataset: A benchmark for multiple-object tracking and counting

    Justin Kay, Peter Kulits, Suzanne Stathatos, Siqi Deng, Erik Young, Sara Beery, Grant Van Horn, and Pietro Perona. The caltech fish counting dataset: A benchmark for multiple-object tracking and counting. InEuropean Conference on Computer Vision, pages 290–311. Springer, 2022

    2022

  24. [32]

    Yolov11: An overview of the key architectural enhancements.arXiv preprint arXiv:2410.17725, 2024

    Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements.arXiv preprint arXiv:2410.17725, 2024

    2024 arXiv

  25. [33]

    Avoiding discrimination through causal reasoning.Advances in neural information processing systems, 30, 2017

    Niki Kilbertus, Mateo Rojas Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing, and Bernhard Schölkopf. Avoiding discrimination through causal reasoning.Advances in neural information processing systems, 30, 2017. 11

    2017

  26. [34]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  27. [35]

    Gaussian-based pooling for convolutional neural networks.Advances in Neural Information Processing Systems, 32, 2019

    Takumi Kobayashi. Gaussian-based pooling for convolutional neural networks.Advances in Neural Information Processing Systems, 32, 2019

    2019

  28. [36]

    Semi-supervised learning, causality, and the conditional cluster assumption

    Julius Kügelgen, Alexander Mey, Marco Loog, and Bernhard Schölkopf. Semi-supervised learning, causality, and the conditional cluster assumption. InConference on uncertainty in artificial intelligence, pages 1–10. PMLR, 2020

    2020

  29. [37]

    Counterfactual fairness

    Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. Counterfactual fairness. Advances in neural information processing systems, 30, 2017

    2017

  30. [38]

    Disentangle and remerge: interventional knowledge distillation for few-shot object detection from a conditional causal perspective

    Jiangmeng Li, Yanan Zhang, Wenwen Qiang, Lingyu Si, Chengbo Jiao, Xiaohui Hu, Changwen Zheng, and Fuchun Sun. Disentangle and remerge: interventional knowledge distillation for few-shot object detection from a conditional causal perspective. Inproceedings of the AAAI Conferenc...

    2023

  31. [39]

    Bridging composite and real: towards end-to-end deep image matting.International Journal of Computer Vision, 130(2):246– 266, 2022

    Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. Bridging composite and real: towards end-to-end deep image matting.International Journal of Computer Vision, 130(2):246– 266, 2022

    2022

  32. [40]

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

    Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5291–5300, 2022

    2022

  33. [41]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2117–2125, 2017

    2017

  34. [42]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InProceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017

    2017

  35. [43]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. InComputer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings,...

    2014

  36. [44]

    A causal debiasing framework for unsupervised salient object detection

    Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1610–1619, 2022

    2022

  37. [45]

    Graph pooling for graph neural networks: Progress, challenges, and opportunities.arXiv preprint arXiv:2204.07321, 2022

    Chuang Liu, Yibing Zhan, Jia Wu, Chang Li, Bo Du, Wenbin Hu, Tongliang Liu, and Dacheng Tao. Graph pooling for graph neural networks: Progress, challenges, and opportunities.arXiv preprint arXiv:2204.07321, 2022

    2022 arXiv

  38. [46]

    Ssd: Single shot multibox detector

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander 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, pa...

    2016

  39. [47]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017

    2017

  40. [48]

    Unified-iou: For high-quality object detection.arXiv preprint arXiv:2408.06636, 2024

    Xiangjie Luo, Zhihao Cai, Bo Shao, and Yingxun Wang. Unified-iou: For high-quality object detection.arXiv preprint arXiv:2408.06636, 2024

    2024 arXiv

  41. [49]

    Implicit self-regularization in deep neural networks: Evidence from random matrix theory and implications for learning.Journal of Machine Learning Research, 22(165):1–73, 2021

    Charles H Martin and Michael W Mahoney. Implicit self-regularization in deep neural networks: Evidence from random matrix theory and implications for learning.Journal of Machine Learning Research, 22(165):1–73, 2021. 12

    2021

  42. [50]

    Predicting trends in the quality of state-of-the-art neural networks without access to training or testing data.Nature Communica- tions, 12(1):4122, 2021

    Charles H Martin, Tongsu Peng, and Michael W Mahoney. Predicting trends in the quality of state-of-the-art neural networks without access to training or testing data.Nature Communica- tions, 12(1):4122, 2021

    2021

  43. [51]

    Foreground activation maps for weakly supervised object localization

    Meng Meng, Tianzhu Zhang, Qi Tian, Yongdong Zhang, and Feng Wu. Foreground activation maps for weakly supervised object localization. InProceedings of the IEEE/CVF international conference on computer vision, pages 3385–3395, 2021

    2021

  44. [52]

    Watershed of a continuous function.Signal processing, 38(1):99–112, 1994

    Laurent Najman and Michel Schmitt. Watershed of a continuous function.Signal processing, 38(1):99–112, 1994

    1994

  45. [53]

    Revisiting self-distillation.arXiv preprint arXiv:2206.08491, 2022

    Minh Pham, Minsu Cho, Ameya Joshi, and Chinmay Hegde. Revisiting self-distillation.arXiv preprint arXiv:2206.08491, 2022

    2022 arXiv

  46. [54]

    Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization

    Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. Inproceedings of the IEEE/CVF winter conference on applications of computer vision, pages 983–991, 2020

    2020

  47. [55]

    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

    2016

  48. [56]

    Causal bert: Improving object detection by searching for challenging groups

    Cinjon Resnick, Or Litany, Amlan Kar, Karsten Kreis, James Lucas, Kyunghyun Cho, and Sanja Fidler. Causal bert: Improving object detection by searching for challenging groups. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2972–2981, 2021

    2021

  49. [57]

    Hierarchical confusion matrix for classification performance evaluation.Journal of the Royal Statistical Society Series C: Applied Statistics, 72(5):1394–1412, 2023

    Kevin Riehl, Michael Neunteufel, and Martin Hemberg. Hierarchical confusion matrix for classification performance evaluation.Journal of the Royal Statistical Society Series C: Applied Statistics, 72(5):1394–1412, 2023

    2023

  50. [58]

    Semantic foggy scene understanding with synthetic data.International Journal of Computer Vision, 126(9):973–992, Sep 2018

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data.International Journal of Computer Vision, 126(9):973–992, Sep 2018

    2018

  51. [59]

    Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634, 2021

    Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634, 2021

    2021

  52. [60]

    Image compositing is all you need for data augmentation.arXiv preprint arXiv:2502.13936, 2025

    Ang Jia Ning Shermaine, Michalis Lazarou, and Tania Stathaki. Image compositing is all you need for data augmentation.arXiv preprint arXiv:2502.13936, 2025

    2025 arXiv

  53. [61]

    Don’t judge an object by its context: learning to overcome contextual bias

    Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don’t judge an object by its context: learning to overcome contextual bias. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11070–...

    2020

  54. [62]

    Quantifying context bias in domain adaptation for object detection

    Hojun Son and Arpan Kusari. Quantifying context bias in domain adaptation for object detection. arXiv preprint arXiv:2409.14679, 2024

    2024 arXiv

  55. [63]

    Adapool: Exponential adaptive pooling for information- retaining downsampling.IEEE Transactions on Image Processing, 32:251–266, 2022

    Alexandros Stergiou and Ronald Poppe. Adapool: Exponential adaptive pooling for information- retaining downsampling.IEEE Transactions on Image Processing, 32:251–266, 2022

    2022

  56. [64]

    Refining activation downsam- pling with softpool

    Alexandros Stergiou, Ronald Poppe, and Grigorios Kalliatakis. Refining activation downsam- pling with softpool. InProceedings of the IEEE/CVF international conference on computer vision, pages 10357–10366, 2021

    2021

  57. [65]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational conference on machine learning, pages 6105–6114. PMLR, 2019

    2019

  58. [66]

    Improving object detection via local-global contrastive learning.arXiv preprint arXiv:2410.05058, 2024

    Danai Triantafyllidou, Sarah Parisot, Ales Leonardis, and Steven McDonagh. Improving object detection via local-global contrastive learning.arXiv preprint arXiv:2410.05058, 2024

    2024 arXiv

  59. [67]

    A realism metric for generated lidar point clouds.International Journal of Computer Vision, 130(12):2962–2979, 2022

    Larissa T Triess, Christoph B Rist, David Peter, and J Marius Zöllner. A realism metric for generated lidar point clouds.International Journal of Computer Vision, 130(12):2962–2979, 2022. 13

    2022

  60. [68]

    On the fairness of causal algorithmic recourse

    Julius V on Kügelgen, Amir-Hossein Karimi, Umang Bhatt, Isabel Valera, Adrian Weller, and Bernhard Schölkopf. On the fairness of causal algorithmic recourse. InProceedings of the AAAI conference on artificial intelligence, volume 36, pages 9584–9594, 2022

    2022

  61. [69]

    A method for constructing a loss function for multi-scale object detection networks.Sensors, 25(6):1738, 2025

    Dong Wang, Hong Zhu, Yue Zhao, and Jing Shi. A method for constructing a loss function for multi-scale object detection networks.Sensors, 25(6):1738, 2025

    2025

  62. [70]

    Causal attention for unbiased visual recognition

    Tan Wang, Chang Zhou, Qianru Sun, and Hanwang Zhang. Causal attention for unbiased visual recognition. InProceedings of the IEEE/CVF international conference on computer vision, pages 3091–3100, 2021

    2021

  63. [71]

    A general deep learning model for bird detection in high-resolution airborne imagery.Ecological Applications, 32(8):e2694, 2022

    Ben G Weinstein, Lindsey Garner, Vienna R Saccomanno, Ashley Steinkraus, Andrew Ortega, Kristen Brush, Glenda Yenni, Ann E McKellar, Rowan Converse, Christopher D Lippitt, et al. A general deep learning model for bird detection in high-resolution airborne imagery.Ecological Ap...

    2022

  64. [72]

    Ben G Weinstein, Sarah J Graves, Sergio Marconi, Aditya Singh, Alina Zare, Dylan Stewart, Stephanie A Bohlman, and Ethan P White. A benchmark dataset for canopy crown detection and delineation in co-registered airborne rgb, lidar and hyperspectral imagery from the national eco...

    2021

  65. [73]

    Background activation suppression for weakly supervised object localization

    Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14228–14237. IEEE, 2022

    2022

  66. [74]

    Discover and cure: Concept- aware mitigation of spurious correlation

    Shirley Wu, Mert Yuksekgonul, Linjun Zhang, and James Zou. Discover and cure: Concept- aware mitigation of spurious correlation. InInternational Conference on Machine Learning, pages 37765–37786. PMLR, 2023

    2023

  67. [75]

    Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains

    Mingjun Xu, Lingyun Qin, Weijie Chen, Shiliang Pu, and Lei Zhang. Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8103–...

    2023

  68. [76]

    Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains

    Mingjun Xu, Lingyun Qin, Weijie Chen, Shiliang Pu, and Lei Zhang. Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8103–8112, 2023

    2023

  69. [77]

    Liftpool: Lifting-based graph pooling for hierarchical graph representation learning.arXiv preprint arXiv:2204.12881, 2022

    Mingxing Xu, Wenrui Dai, Chenglin Li, Junni Zou, and Hongkai Xiong. Liftpool: Lifting-based graph pooling for hierarchical graph representation learning.arXiv preprint arXiv:2204.12881, 2022

    2022 arXiv

  70. [78]

    Cross dataset analysis of domain shift in cxr lung region detection.Diagnostics, 13(6):1068, 2023

    Zhiyun Xue, Feng Yang, Sivaramakrishnan Rajaraman, Ghada Zamzmi, and Sameer Antani. Cross dataset analysis of domain shift in cxr lung region detection.Diagnostics, 13(6):1068, 2023

    2023

  71. [79]

    Stochastic pooling for regularization of deep convolutional neural networks.arXiv preprint arXiv:1301.3557, 2013

    Matthew D Zeiler and Rob Fergus. Stochastic pooling for regularization of deep convolutional neural networks.arXiv preprint arXiv:1301.3557, 2013

    2013 arXiv

  72. [80]

    S3pool: Pooling with stochastic spatial sampling

    Shuangfei Zhai, Hui Wu, Abhishek Kumar, Yu Cheng, Yongxi Lu, Zhongfei Zhang, and Rogerio Feris. S3pool: Pooling with stochastic spatial sampling. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4970–4978, 2017

    2017

  73. [81]

    Background activation suppression for weakly supervised object localization and semantic segmentation.International Journal of Computer Vision, 132(3):750–775, 2024

    Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation.International Journal of Computer Vision, 132(3):750–775, 2024

    2024

  74. [82]

    Fairness in decision-making—the causal explanation formula

    Junzhe Zhang and Elias Bareinboim. Fairness in decision-making—the causal explanation formula. InProceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018

    2018

  75. [83]

    Causal reasoning in typical computer vision tasks.Science China Technological Sciences, 67(1):105–120, 2024

    Kexuan Zhang, Qiyu Sun, Chaoqiang Zhao, and Yang Tang. Causal reasoning in typical computer vision tasks.Science China Technological Sciences, 67(1):105–120, 2024. 14

    2024

  76. [84]

    Adapting object detectors via selective cross-domain alignment

    Xinge Zhu, Jiangmiao Pang, Ceyuan Yang, Jianping Shi, and Dahua Lin. Adapting object detectors via selective cross-domain alignment. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 687–696, 2019

    2019

  77. [85]

    Learning data augmentation strategies for object detection

    Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16, pages 566–583...

    2020

Pith tools

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