Pith. sign in

REVIEW 4 major objections 5 minor 81 references

A Reverse Causal Framework to Mitigate Spurious Correlations for Debiasing Scene Graph Generation

T0 review · 4 major / 5 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 recasting two-stage scene graph training as a reverse causal chain removes spurious correlations and eliminates both head-tail and foreground-background biases.

desk verdict Solid empirical debiasing method with a broken causal story; the method works, but not for the reasons the paper gives. read the letter →

arxiv 2505.23451 v1 pith:3J5AEX3B submitted 2025-05-29 cs.CV

Shuzhou Sun , Li Liu , Tianpeng Liu , Shuaifeng Zhi , Ming-Ming Cheng , Janne Heikkilä , Yongxiang Liu This is my paper
classification cs.CV
keywords scenegraphgenerationspuriouscorrelationcausalinferencedebiasinglong-taileddistributionactivelearningmeanrecallforeground-backgroundbias
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

Two-stage scene graph generation trains a frozen detector to extract relationship features and then a classifier to label them, which the paper models as the causal chain $X \to R \to Y$. The paper argues that because features are aggregated per image rather than sampled independently, this chain makes predictions for one relationship depend on the other relationships in the scene, producing head-tail bias and foreground-background bias. To remove those biases it redraws the structure as a reverse causal graph $X \to R \leftarrow Y$, treating the relationship features as a confounder to be intervened on. A loss-driven Active Reverse Estimation scheme resamples relationship features toward category balance, and Maximum Information Sampling diversifies object pairs. The paper claims the trained classifier is Bayes optimal and reports state-of-the-art mean recall on VG150, GQA, Open Images V6, and PSG.

What carries the argument

The load-bearing object is the reverse causal structure $X \to R \leftarrow Y$, in which the relationship feature vector $R$ is a collider that separates the image $X$ from the prediction $Y$. Active Reverse Estimation (ARE) operationalizes the reverse edge $R \leftarrow Y$: it derives a sampling distribution from the exponential of the negative per-category loss of the previous batch, then samples those categories into the current batch without replacement, so the feature space is refined by training outcomes rather than by the raw image distribution. Maximum Information Sampling (MIS) replaces naive random sampling inside ARE by preferring object pairs that add the most information, countering object-pair imbalance. The background-to-foreground ratio $\pi$ completes the intervention by keeping background relationships from swamping the classifier.

What would settle it

Simulate the paper's graph with a small dataset in which the image $X$ determines features $R$ and the label $Y$ also determines $R$; restrict to samples with a fixed value of $R$ and check whether $X$ and $Y$ remain independent—standard collider behavior makes them dependent, while the paper's Eq. (4) predicts independence.

Watch

Extended reading notes

Core claim

The central claim is that bias in two-stage scene graph generation is not only a data-imbalance problem but a structural artifact of the training pipeline. The paper asserts that rewriting the pipeline $X \to R \to Y$ as $X \to R \leftarrow Y$ makes the image $X$ and the predictions $Y$ independent, so the relationship features sampled during training behave like independent draws and the classifier reaches the Bayes optimum. This is implemented through Active Reverse Estimation, which uses the previous batch's per-category losses as query conditions to sample relationship features into the current batch, and Maximum Information Sampling, which prefers diverse object pairs. Background relationships are clipped to a fixed ratio with foreground ones. The paper reports that this removes both the head-tail bias and the previously understudied foreground-background bias, with the largest measured gains in mean recall, and that the same recipe transfers across framework backbones and even to one-stage scene graph models.

Load-bearing premise

The argument depends on the premise that redrawing the training pipeline as image → features ← prediction makes the image and the prediction independent, even though the model always has the features in hand during training.

Editorial extensions

If this is right

  • Head-tail bias is reduced: tail relationships such as 'sitting on' and 'standing on' are recalled far more often than under the backbone, without collapsing overall recall on head relationships.
  • Foreground-background bias is reduced: when background outputs are included in evaluation, RcSGG retains a large fraction of its mean recall, whereas baselines drop to near zero.
  • The method is model-agnostic: it improves mean recall when plugged into MotifsNet, VCTree, and Transformer backbones, and also lifts mean recall in end-to-end SGG frameworks.
  • Testing costs nothing extra: the intervention happens only during training, and inference follows the standard $X \to R \to Y$ flow.
  • Approximate batch balance is the formal target: Theorem 3 states that per-batch category counts that are merely close to balanced suffice for the Bayes optimal classifier.

Reading between the lines

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

  • A reader could separate the sampling recipe from the causal story: the measurable gains may come from loss-weighted oversampling and background clipping, and the claim that conditioning on a collider preserves independence is not supported by standard causal-graph theory.
  • A testable extension is to run ARE and MIS with the original chain $X \to R \to Y$ left intact; if mean recall gains persist, the active sampling scheme is the active ingredient, not the reversed graph.
  • The framework suggests a cheap robustness probe for any debiasing method: evaluate with background relationships included, since current protocols mask how much models lean on the background class.
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 / 5 minor

Summary. The paper presents RcSGG, a training-time intervention framework for debiased scene graph generation. It argues that two-stage SGG pipelines X→R→Y create spurious correlations between images and relationship predictions, producing head-tail and foreground-background biases. The method reconstructs the causal graph as X→R←Y, samples relationship features with Active Reverse Estimation (ARE) driven by previous-batch losses, and uses Maximum Information Sampling (MIS) to diversify object pairs while retaining a π:1 background-to-foreground ratio. Experiments are reported on VG150, GQA, Open Images V6, and PSG with MotifsNet, VCTree, and Transformer backbones, showing consistent mean-recall improvements.

Significance. If the causal claims held, RcSGG would provide a general, model-agnostic debiasing principle for SGG with a theoretical guarantee. The empirical study is a clear strength: four datasets, three backbones, three evaluation modes, background-inclusive evaluation, component and hyperparameter ablations, efficiency measurements, and extension to one-stage and panoptic SGG. However, the central theoretical guarantee is not supported as stated: the D-separation argument in §3.3 is incorrect, and the Bayes-optimality theorems are either conditional on unverified assumptions or rest on an invalid convergence argument. The method may still be valuable as a feature-space rebalancing and diversification technique, but its significance as a causal-inference contribution is currently overstated.

major comments (4)
  1. [§3.3, Eq. (4)] The paper claims that after replacing X→R→Y by X→R←Y, X and Y are independent, 'even when R is given,' because the paths from X→R and R←Y are decoupled. This contradicts standard D-separation: in X→R←Y, R is a collider, and conditioning on R opens the path between X and Y, inducing dependence (Berkson's paradox). Equation (4) only marginalizes over R and does not prove P(X,Y|R)=P(X|R)P(Y|R). Since the classifier is trained and evaluated on R, the conditional distribution given R is the relevant quantity, so the reverse causal structure does not by itself eliminate the spurious correlations the paper targets. Theorems 1–3 are balanced-batch results and are independent of the DAG reversal, so they do not repair this gap.
  2. [§3.2, Assumption 1 and §3.3.1, Theorem 3] The paper defines spurious correlations as within-batch predictive dependence (Assumption 1) and then concludes that RcSGG removes them because ARE achieves approximate batch balance, which Theorem 3 equates with Bayes optimality. This reasoning is close to circular: the conclusion is largely built into the definition of spurious correlations and into the sampling objective, which is explicitly constructed to balance categories. No argument links the reverse-causal graph to the balanced-batch property, and the hyperparameters used in the experiments (K′=44, π=3, λ=0.01) are not shown to satisfy the approximate balance condition |r_bm|≈|r_bn|. The observed gains may therefore be attributable to rebalancing and feature diversification rather than to an intervention on a confounder.
  3. [Appendix E, Theorem 3 and Eq. (29)–(30)] The proof of Theorem 3 is not valid. Equation (29) asserts a uniform bound δ on the deviation between batch gradients without justification, and Equation (30) shows only that the accumulated deviation after n iterations is at most nδ, which grows with n; it does not imply that the iterates converge to f*_c. Additionally, the 'ideal training condition' footnote to Appendix D assumes the optimizer reaches a global optimum, so Theorem 1 and its descendants are conditional on an assumption that is not verified and is generally false for non-convex deep networks. The claim that fc is a Bayesian optimal classifier is therefore unsupported as stated.
  4. [Appendix A and Figure 4] Appendix A introduces an unobserved confounder Z with X=a1Z+ε1 and Y=a2Z+ε2 to prove that spurious correlations exist, but the SCM in Figure 4 does not include Z. If such a confounder exists, the graph X→R←Y is not a complete causal model, and X and Y need not be independent even marginally. The correlation computation in Eqs. (15)–(17) also assumes the linear confounding structure it aims to establish; the empirical observation of relationship co-occurrence does not imply that specific generative form. This weakens the theoretical motivation for both the spurious-correlation analysis and the reverse-causal reconstruction.
minor comments (5)
  1. [§3.3] The text 'the relationship classifier fo in the standard two-stage SGG framework' should read fc, since fo denotes the object detector; this notation error occurs in the sentence introducing the SCM.
  2. [§4.1] The text says experiments were conducted across 'three datasets, VG150, GQA, Open Images V6, and PSG,' but four datasets are listed; the count should be corrected.
  3. [Figure 3(b)] The axis label 'Sheet1' and 'X轴标题' are spreadsheet artifacts and should be replaced with a proper caption and axis labels.
  4. [Table 3] The definitions of MR@KΔ and MR@K♢ should specify the exact averages being combined; the current footnote is ambiguous because AVGΔR and AVG♢R already combine three or two K values.
  5. [References] References [54] and [76] appear to duplicate the same Structured Sparse R-CNN paper; the duplicate should be removed or cross-referenced.

Circularity Check

2 steps flagged · score 6.0 of 10

The theoretical guarantee is largely self-definitional: spurious correlations are defined as batch-level imbalance, ARE is built to approximate batch balance, and Theorem 3 then certifies Bayes optimality; the reverse-causal independence argument in Eq. 4 also assumes the independence it claims to prove.

  1. self definitional [Section 3.3, Equation (4) and the surrounding D-separation argument]
    "According to the D-Separation principle, if R is not given, X and Y are independent since R acts as a collider, resulting in P (X, Y) = P (X)P (Y ). Additionally, even when R is given, the paths from X → R and R ← Y are decoupled, preventing collider bias and maintaining independence between X and Y. Hence, we still have: P (X, Y) = P (X)P (Y ) Σ_R P (R | X, Y) = P (X)P (Y )."

    The display writes the joint as P(X,R,Y)=P(X)P(Y)P(R|X,Y). That factorization already contains the marginal independence P(X,Y)=P(X)P(Y) that the equation is supposed to establish; the correct expansion would be P(X,R,Y)=P(R|X,Y)P(X,Y). Moreover, the classifier is trained on R, so the operative requirement is conditional independence given R, and the sentence 'even when R is given ... maintaining independence between X and Y' simply stipulates that required property. The claimed elimination of spurious correlations is thus an assumption baked into the reverse-causal construction, not a derived result.

  2. self definitional [Section 3.2 (Assumptions 1–2) and Section 3.3 (ARE and Theorem 3)]
    "Assumptions 1 and 2 highlight the sources of spurious correlations in the SGG task, thereby revealing that we only need to balance within each batch to prevent spurious correlations. ... This enables ARE to optimize the relationship feature R effectively, leading to an approximate batch balance. While this may not fully satisfy the strict balance conditions of Theorem 2, it still supports achieving a Bayesian optimal classifier as outlined in Theorem 3."

    Spurious correlations are operationally defined as within-batch inter-feature influence in Assumption 1. The paper then states that batch balance 'prevent[s] spurious correlations,' and ARE is explicitly designed to achieve approximate batch balance. Theorem 3 certifies that approximate batch balance yields a Bayes-optimal classifier, which the paper equates with being unaffected by spurious correlations. Consequently, the headline claim that RcSGG 'can mitigate the spurious correlations inherent in the SGG framework' follows directly from the paper's own definition of spurious correlations and the construction of ARE; the reverse-causal DAG is not providing independent theoretical support. The conclusion is built into the definitions.

full rationale

The empirical portion of the paper is self-contained: RcSGG is evaluated against external benchmarks and compared with many published methods, and the reported mR@K improvements do not depend on circular reasoning. The self-citations ([37], [51]) are contextual or comparative and are not load-bearing proof, so they do not raise the score by themselves. However, the central theoretical claim is partially circular. First, the paper defines spurious correlations as within-batch dependence and then equates removing them with achieving batch balance; ARE is a batch-balancing sampler, and Theorem 3 simply certifies approximate batch balance, making the 'mitigation of spurious correlations' an artifact of the definition. Second, the reverse-causal derivation in Eq. 4 factors the joint distribution in a way that already assumes X and Y are independent, and the paper further stipulates the conditional independence given R that the trained classifier would actually need. These are not merely missing steps; the conclusion is contained in the premises. There are also non-circular correctness concerns, notably the false claim that conditioning on a collider preserves independence, but the circularity score is driven by the definitional reduction described above. A score of 6 reflects partial circularity in the theoretical guarantee while acknowledging that the method is a concrete, benchmark-tested sampling scheme whose empirical gains may be real.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim rests on several assumptions that are either stated as postulates (Assumptions 1 and 2), assumed domain structure (X → R → Y), or ideal training conditions. Four hyperparameters are fitted to validation data. The invented confounder Z is not used in the method and has no falsifiable predictions.

free parameters (4)
  • π (foreground-background ratio) = 3.0
    Chosen by one-by-one validation search (Table 6); controls the ratio of background to foreground relationships retained per batch.
  • K' (query set size) = 44
    Number of relationship categories in the query set, chosen by validation (Table 7); optimal when selected instances cover about 20% of the data.
  • α (loss-weight factor) = 0.2
    Controls the sharpness of the loss-based sampling distribution (Eq. 6); tuned on validation (Table 8).
  • λ (sampling magnitude) = 0.01
    Scales the number of samples per category; set by hand based on the order of magnitude of K'.
assumptions (5)
  • ad hoc to paper Assumption 1: relationships within the same batch influence each other's predictions.
    Stated in Section 3.2 as the definition of spurious correlation; not derived from data or theory.
  • ad hoc to paper Assumption 2: no spurious correlations across batches.
    Needed for the batch-independence claim; violated in practice because minibatches are drawn from the same dataset and share features.
  • domain assumption The two-stage SGG pipeline can be modeled as the causal chain X → R → Y.
    Used in Section 3.3 to justify the reverse causal structure; treated as given.
  • ad hoc to paper Ideal training condition: the optimizer reaches the global optimum.
    Invoked in Appendix D to prove Theorem 1; not guaranteed for neural network training.
  • ad hoc to paper An unobserved confounder Z accounts for the correlation between X and Y (Appendix A).
    Postulated to demonstrate spurious correlations; not measured or used in the method.
invented entities (1)
  • Unobserved confounder Z
    purpose: To prove that X and Y are spuriously correlated in the two-stage framework
    Defined in Appendix A with X = a1 Z + ε1, Y = a2 Z + ε2. No direct or falsifiable handle is provided; it is a narrative device.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Reverse Causal Framework to Mitigate Spurious Correlations for Debiasing Scene Graph Generation." pith.science (2026). https://pith.science/paper/3J5AEX3B

@misc{pith2026250523451,
  author       = {Pith},
  title        = {Pith review of: A Reverse Causal Framework to Mitigate Spurious Correlations for Debiasing Scene Graph Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3J5AEX3B}},
  note         = {Machine review of arXiv:2505.23451}
}
read the original abstract

Existing two-stage Scene Graph Generation (SGG) frameworks typically incorporate a detector to extract relationship features and a classifier to categorize these relationships; therefore, the training paradigm follows a causal chain structure, where the detector's inputs determine the classifier's inputs, which in turn influence the final predictions. However, such a causal chain structure can yield spurious correlations between the detector's inputs and the final predictions, i.e., the prediction of a certain relationship may be influenced by other relationships. This influence can induce at least two observable biases: tail relationships are predicted as head ones, and foreground relationships are predicted as background ones; notably, the latter bias is seldom discussed in the literature. To address this issue, we propose reconstructing the causal chain structure into a reverse causal structure, wherein the classifier's inputs are treated as the confounder, and both the detector's inputs and the final predictions are viewed as causal variables. Specifically, we term the reconstructed causal paradigm as the Reverse causal Framework for SGG (RcSGG). RcSGG initially employs the proposed Active Reverse Estimation (ARE) to intervene on the confounder to estimate the reverse causality, \ie the causality from final predictions to the classifier's inputs. Then, the Maximum Information Sampling (MIS) is suggested to enhance the reverse causality estimation further by considering the relationship information. Theoretically, RcSGG can mitigate the spurious correlations inherent in the SGG framework, subsequently eliminating the induced biases. Comprehensive experiments on popular benchmarks and diverse SGG frameworks show the state-of-the-art mean recall rate.

Figures

Figures reproduced from arXiv: 2505.23451 by the authors.

Figure 1
Figure 1. The motivations of RcSGG. Images and count numbers are sourced from VG150. Test results were obtained from the MotifsNet model under the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The standard pipeline of the two-stage scene graph generation framework, where the input image [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (a) Sampling independence of the relationship feature space and the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Structural Causal Model (SCM). (a) is the causal chain structure of the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: (a) Distribution of relationship features across the entire dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Given the distribution of object pairs under specific relationships, for [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Recall rate (R@100) for representative head and tail relationships, [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results of our method and baseline approach (MotifsNet [ [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Average output logits. in the baseline method overwhelmingly dominate, weakening the model’s response to foreground relationships (almost all negative). In contrast, our proposed approach significantly improves this unhealthy logit distribution, enhancing the model’s r…
Figure 10
Figure 10. Figure 10: Compared with classical uncertainty-driven active learning [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: The statistics of our method’s wrong predictions on the test set. In [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 77 canonical work pages

  1. [1]

    Image retrieval using scene graphs,

    J. Johnson, R. Krishna, M. Stark, L.-J. Li, D. Shamma, M. Bernstein, and L. Fei-Fei, “Image retrieval using scene graphs,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2015, pp. 3668–3678. IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 17

    2015

  2. [2]

    A comprehensive survey of scene graphs: Generation and application,

    X. Chang, P . Ren, P . Xu, Z. Li, X. Chen, and A. Hauptmann, “A comprehensive survey of scene graphs: Generation and application,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 1, pp. 1–26, 2023

    work page 2023

  3. [3]

    Visual commonsense r-cnn,

    T. Wang, J. Huang, H. Zhang, and Q. Sun, “Visual commonsense r-cnn,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2020, pp. 10 760–10 770

    work page 2020

  4. [4]

    Scene graph generation with role-playing large language models,

    G. Chen, J. Li, and W. Wang, “Scene graph generation with role-playing large language models,” Int. Conf. Neural Inf. Process. Syst. , 2024

    work page 2024

  5. [5]

    Neural motifs: Scene graph parsing with global context,

    R. Zellers, M. Yatskar, S. Thomson, and Y. Choi, “Neural motifs: Scene graph parsing with global context,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 5831–5840

    work page 2018

  6. [6]

    Graph r-cnn for scene graph generation,

    J. Yang, J. Lu, S. Lee, D. Batra, and D. Parikh, “Graph r-cnn for scene graph generation,” in Proc. Eur. Conf. Comput. Vis., 2018, pp. 670–685

    work page 2018

  7. [7]

    Scene graph generation by iterative message passing,

    D. Xu, Y. Zhu, C. B. Choy, and L. Fei-Fei, “Scene graph generation by iterative message passing,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2017, pp. 5410–5419

    work page 2017

  8. [8]

    Learning semantic relationships for better action retrieval in images,

    V . Ramanathan, C. Li, J. Deng, W. Han, Z. Li, K. Gu, Y. Song, S. Bengio, C. Rosenberg, and L. Fei-Fei, “Learning semantic relationships for better action retrieval in images,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2015, pp. 1100–1109

    work page 2015

Show all 81 references
  1. [9]

    Learning conditioned graph structures for interpretable visual question answering,

    W. Norcliffe-Brown, S. Vafeias, and S. Parisot, “Learning conditioned graph structures for interpretable visual question answering,” in Proc. 32nd Int. Conf. Neural Inf. Process. Syst. , vol. 31, 2018

    2018

  2. [10]

    Explainable and explicit visual reasoning over scene graphs,

    J. Shi, H. Zhang, and J. Li, “Explainable and explicit visual reasoning over scene graphs,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2019, pp. 8376–8384

    2019

  3. [11]

    Auto-encoding scene graphs for image captioning,

    X. Yang, K. Tang, H. Zhang, and J. Cai, “Auto-encoding scene graphs for image captioning,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2019, pp. 10 685–10 694

    2019

  4. [12]

    Unpaired image captioning via scene graph alignments,

    J. Gu, S. Joty, J. Cai, H. Zhao, X. Yang, and G. Wang, “Unpaired image captioning via scene graph alignments,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2019, pp. 10 323–10 332

    2019

  5. [13]

    Segmentation-grounded scene graph generation,

    S. Khandelwal, M. Suhail, and L. Sigal, “Segmentation-grounded scene graph generation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. , 2021, pp. 15 879–15 889

    2021

  6. [14]

    Fine- grained scene graph generation with data transfer,

    A. Zhang, Y. Yao, Q. Chen, W. Ji, Z. Liu, M. Sun, and T.-S. Chua, “Fine- grained scene graph generation with data transfer,” in Proc. Eur. Conf. Comput. Vis., 2022

    2022

  7. [15]

    Fine-grained predicates learning for scene graph generation,

    X. Lyu, L. Gao, Y. Guo, Z. Zhao, H. Huang, H. T. Shen, and J. Song, “Fine-grained predicates learning for scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2022, pp. 19 467–19 475

    2022

  8. [16]

    PPDL: Predicate probability distribution based loss for unbiased scene graph generation,

    W. Li, H. Zhang, Q. Bai, G. Zhao, N. Jiang, and X. Yuan, “PPDL: Predicate probability distribution based loss for unbiased scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2022, pp. 19 447–19 456

    2022

  9. [17]

    Stacked hybrid- attention and group collaborative learning for unbiased scene graph generation,

    X. Dong, T. Gan, X. Song, J. Wu, Y. Cheng, and L. Nie, “Stacked hybrid- attention and group collaborative learning for unbiased scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2022, pp. 19 427–19 436

    2022

  10. [18]

    Energy-based learning for scene graph generation,

    M. Suhail, A. Mittal, B. Siddiquie, C. Broaddus, J. Eledath, G. Medioni, and L. Sigal, “Energy-based learning for scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2021, pp. 13 936–13 945

    2021

  11. [19]

    Recovering the unbiased scene graphs from the biased ones,

    M.-J. Chiou, H. Ding, H. Yan, C. Wang, R. Zimmermann, and J. Feng, “Recovering the unbiased scene graphs from the biased ones,” in Proc. 29th ACM Int. Conf. Multi., 2021, pp. 1581–1590

    2021

  12. [20]

    Unbiased scene graph generation from biased training,

    K. Tang, Y. Niu, J. Huang, J. Shi, and H. Zhang, “Unbiased scene graph generation from biased training,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2020, pp. 3716–3725

    2020

  13. [21]

    Hierarchical memory learning for fine-grained scene graph generation,

    Y. Deng, Y. Li, Y. Zhang, X. Xiang, J. Wang, J. Chen, and J. Ma, “Hierarchical memory learning for fine-grained scene graph generation,” in Proc. Eur. Conf. Comput. Vis., 2022

    2022

  14. [22]

    The devil is in the labels: Noisy label correction for robust scene graph generation,

    L. Li, L. Chen, Y. Huang, Z. Zhang, S. Zhang, and J. Xiao, “The devil is in the labels: Noisy label correction for robust scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2022, pp. 18 869–18 878

    2022

  15. [23]

    Resistance training using prior bias: toward unbiased scene graph generation,

    C. Chen, Y. Zhan, B. Yu, L. Liu, Y. Luo, and B. Du, “Resistance training using prior bias: toward unbiased scene graph generation,” in Proc. AAAI Conf. Artif. Intell., 2022

    2022

  16. [24]

    From general to specific: Informative scene graph generation via balance adjustment,

    Y. Guo, L. Gao, X. Wang, Y. Hu, X. Xu, X. Lu, H. T. Shen, and J. Song, “From general to specific: Informative scene graph generation via balance adjustment,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 16 383–16 392

    2021

  17. [25]

    Learning to compose dynamic tree structures for visual contexts,

    K. Tang, H. Zhang, B. Wu, W. Luo, and W. Liu, “Learning to compose dynamic tree structures for visual contexts,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2019, pp. 6619–6628

    2019

  18. [26]

    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,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 6, pp. 1137–1149, 2017

    2017

  19. [27]

    Causality,

    Pearl, Judea, “Causality,” Cambridge university press, 2009

    2009

  20. [28]

    A causal framework for distribution generalization,

    R. Christiansen, N. Pfister, M. E. Jakobsen, N. Gnecco, and J. Peters, “A causal framework for distribution generalization,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 10, pp. 6614–6630, 2022

    2022

  21. [29]

    Causal diagrams for empirical research,

    J. Pearl, “Causal diagrams for empirical research,” Biometrika, vol. 82, no. 4, pp. 669–688, 1995

    1995

  22. [30]

    Models, reasoning and inference,

    J. Pearl et al. , “Models, reasoning and inference,” Cambridge, UK: CambridgeUniversityPress, vol. 19, no. 2, 2000

    2000

  23. [31]

    When causal inference meets deep learning,

    Y. Luo, J. Peng, and J. Ma, “When causal inference meets deep learning,” Nat. Mach. Intell., vol. 2, no. 8, pp. 426–427, 2020

    2020

  24. [32]

    Toward causal representation learning,

    B. Sch ¨olkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y. Bengio, “Toward causal representation learning,” Proc. of the IEEE, vol. 109, no. 5, pp. 612–634, 2021

    2021

  25. [33]

    Causal fairness analysis,

    D. Plecko and E. Bareinboim, “Causal fairness analysis,” arXiv:2207.11385, 2022

    2022 arXiv

  26. [34]

    Towards fewer annotations: Active learning via region impurity and prediction uncertainty for domain adaptive semantic segmentation,

    B. Xie, L. Yuan, S. Li, C. H. Liu, and X. Cheng, “Towards fewer annotations: Active learning via region impurity and prediction uncertainty for domain adaptive semantic segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2022, pp. 8068–8078

    2022

  27. [35]

    Active learning polynomial threshold functions,

    O. Ben-Eliezer, M. Hopkins, C. Yang, and H. Yu, “Active learning polynomial threshold functions,” Int. Conf. Neural Inf. Process. Syst. , 2022

    2022

  28. [36]

    Learning active learning from data,

    K. Konyushkova, R. Sznitman, and P . Fua, “Learning active learning from data,” Int. Conf. Neural Inf. Process. Syst. , vol. 30, 2017

    2017

  29. [37]

    Evidential uncertainty and diversity guided active learning for scene graph generation,

    S. Sun, S. Zhi, J. Heikkil ¨a, and L. Liu, “Evidential uncertainty and diversity guided active learning for scene graph generation,” in Proc. Int. Conf. Learn. Represent., 2023

    2023

  30. [38]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Int. Conf. Neural Inf. Process. Syst., vol. 25, 2012

    2012

  31. [39]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in Proc. Int. Conf. Learn. Represent., 2014

    2014

  32. [40]

    Visual genome: Connecting language and vision using crowdsourced dense image annotations,

    R. Krishna, Y. Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz, S. Chen, Y. Kalantidis, L.-J. Li, D. A. Shamma et al., “Visual genome: Connecting language and vision using crowdsourced dense image annotations,” Int. J. Comput. Vis., vol. 123, no. 1, pp. 32–73, 2017

    2017

  33. [41]

    Learning of visual relations: The devil is in the tails,

    A. Desai, T.-Y. Wu, S. Tripathi, and N. Vasconcelos, “Learning of visual relations: The devil is in the tails,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 15 404–15 413

    2021

  34. [42]

    Multi-label meta weighting for long-tailed dynamic scene graph generation,

    S. Chen, Y. Du, P . Mettes, and C. G. Snoek, “Multi-label meta weighting for long-tailed dynamic scene graph generation,” in Proc. ACM International Conference on Multimedia Retrieval , 2023, pp. 39–47

    2023

  35. [43]

    Prototype-based embedding network for scene graph generation,

    C. Zheng, X. Lyu, L. Gao, B. Dai, and J. Song, “Prototype-based embedding network for scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2023, pp. 22 783–22 792

    2023

  36. [44]

    Dual- branch hybrid learning network for unbiased scene graph generation,

    C. Zheng, L. Gao, X. Lyu, P . Zeng, A. El Saddik, and H. T. Shen, “Dual- branch hybrid learning network for unbiased scene graph generation,” IEEE Trans. Circuits Syst. Video Technol., 2023

    2023

  37. [45]

    Vision relation transformer for unbiased scene graph generation,

    G. Sudhakaran, D. S. Dhami, K. Kersting, and S. Roth, “Vision relation transformer for unbiased scene graph generation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 21 882–21 893

    2023

  38. [46]

    Visually-prompted language model for fine-grained scene graph generation in an open world,

    Q. Yu, J. Li, Y. Wu, S. Tang, W. Ji, and Y. Zhuang, “Visually-prompted language model for fine-grained scene graph generation in an open world,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. , 2023, pp. 21 560– 21 571

    2023

  39. [47]

    Addressing predicate overlap in scene graph generation with semantic granularity controller,

    G. Chen, L. Li, Y. Luo, and J. Xiao, “Addressing predicate overlap in scene graph generation with semantic granularity controller,” in Proc. IEEE Int. Conf. Multimedia Expo. IEEE, 2023, pp. 78–83

    2023

  40. [48]

    Causal balancing for domain generalization,

    X. Wang, M. Saxon, J. Li, H. Zhang, K. Zhang, and W. Y. Wang, “Causal balancing for domain generalization,” Proc. Int. Conf. Learn. Represent. , 2023

    2023

  41. [49]

    Debiasing graph neural networks via learning disentangled causal substructure,

    S. Fan, X. Wang, Y. Mo, C. Shi, and J. Tang, “Debiasing graph neural networks via learning disentangled causal substructure,” Int. Conf. Neural Inf. Process. Syst., 2022

    2022

  42. [50]

    Debiased scene graph generation for dual imbalance learning,

    H. Zhou, J. Zhang, T. Luo, Y. Yang, and J. Lei, “Debiased scene graph generation for dual imbalance learning,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 4, pp. 4274–4288, 2023

    2023

  43. [51]

    Unbiased scene graph generation via two-stage causal modeling,

    S. Sun, S. Zhi, Q. Liao, J. Heikkil ¨a, and L. Liu, “Unbiased scene graph generation via two-stage causal modeling,” IEEE Trans. Pattern Anal. Mach. Intell., 2023

    2023

  44. [52]

    Bayes imbalance impact index: A measure of class imbalanced data set for classification problem,

    Y. Lu, Y.-M. Cheung, and Y. Y. Tang, “Bayes imbalance impact index: A measure of class imbalanced data set for classification problem,” IEEE Trans Neural Netw Learn Syst., vol. 31, no. 9, pp. 3525–3539, 2019

    2019

  45. [53]

    On exploring undetermined relationships for visual relationship detection,

    Y. Zhan, J. Yu, T. Yu, and D. Tao, “On exploring undetermined relationships for visual relationship detection,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2019, pp. 5128–5137

    2019

  46. [55]

    Unbiased heterogeneous scene graph generation with relation-aware message passing neural network,

    K. Yoon, K. Kim, J. Moon, and C. Park, “Unbiased heterogeneous scene graph generation with relation-aware message passing neural network,” in Proc. AAAI Conf. Artif. Intell., vol. 37, no. 3, 2023, pp. 3285– 3294. IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 18

    2023

  47. [56]

    Devil’s on the edges: Selective quad attention for scene graph generation,

    D. Jung, S. Kim, W. H. Kim, and M. Cho, “Devil’s on the edges: Selective quad attention for scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2023, pp. 18 664–18 674

    2023

  48. [57]

    Fast contextual scene graph generation with unbiased context augmentation,

    T. Jin, F. Guo, Q. Meng, S. Zhu, X. Xi, W. Wang, Z. Mu, and W. Song, “Fast contextual scene graph generation with unbiased context augmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2023, pp. 6302–6311

    2023

  49. [58]

    Adaptive fine-grained predicates learning for scene graph generation,

    X. Lyu, L. Gao, P . Zeng, H. T. Shen, and J. Song, “Adaptive fine-grained predicates learning for scene graph generation,” IEEE Trans. Pattern Anal. Mach. Intell., 2023

    2023

  50. [59]

    Environment-invariant curriculum relation learning for fine-grained scene graph generation,

    Y. Min, A. Wu, and C. Deng, “Environment-invariant curriculum relation learning for fine-grained scene graph generation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 13 296–13 307

    2023

  51. [60]

    Peer learning for unbiased scene graph generation,

    L. Zhou, J. Hu, Y. Zhou, T. L. Lam, and Y. Xu, “Peer learning for unbiased scene graph generation,” arXiv:2301.00146, 2023

    2023 arXiv

  52. [61]

    Compositional feature augmentation for unbiased scene graph generation,

    L. Li, G. Chen, J. Xiao, Y. Yang, C. Wang, and L. Chen, “Compositional feature augmentation for unbiased scene graph generation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 21 685–21 695

    2023

  53. [62]

    Hilo: Exploiting high low frequency relations for unbiased panoptic scene graph generation,

    Z. Zhou, M. Shi, and H. Caesar, “Hilo: Exploiting high low frequency relations for unbiased panoptic scene graph generation,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 21 637–21 648

    2023

  54. [63]

    Nicest: Noisy label correction and training for robust scene graph generation,

    L. Li, L. Chen, H. Shi, H. Zhang, Y. Yang, W. Liu, and J. Xiao, “Nicest: Noisy label correction and training for robust scene graph generation,” arXiv:2207.13316, 2022

    2022 arXiv

  55. [64]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Int. Conf. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  56. [65]

    Gqa: A new dataset for real-world visual reasoning and compositional question answering,

    D. A. Hudson and C. D. Manning, “Gqa: A new dataset for real-world visual reasoning and compositional question answering,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2019, pp. 6700–6709

    2019

  57. [66]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale,

    A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov et al., “The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale,” Int. J. Comput. Vis. , vol...

    1956

  58. [67]

    Bipartite graph network with adaptive message passing for unbiased scene graph generation,

    R. Li, S. Zhang, B. Wan, and X. He, “Bipartite graph network with adaptive message passing for unbiased scene graph generation,” in roc. IEEE Conf. Comput. Vis. Pattern Recognit., 2021, pp. 11 109–11 119

    2021

  59. [68]

    Panoptic scene graph generation,

    J. Yang, Y. Z. Ang, Z. Guo, K. Zhou, W. Zhang, and Z. Liu, “Panoptic scene graph generation,” in Proc. Eur. Conf. Comput. Vis., 2022, pp. 178– 196

    2022

  60. [69]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 770–778

    2016

  61. [70]

    Predicate correlation learning for scene graph generation,

    L. Tao, L. Mi, N. Li, X. Cheng, Y. Hu, and Z. Chen, “Predicate correlation learning for scene graph generation,” IEEE Trans. Image Process., vol. 31, pp. 4173–4185, 2022

    2022

  62. [71]

    Ru-net: Regularized unrolling network for scene graph generation,

    X. Lin, C. Ding, J. Zhang, Y. Zhan, and D. Tao, “Ru-net: Regularized unrolling network for scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2022, pp. 19 457–19 466

    2022

  63. [72]

    Deep active learning for named entity recognition,

    Y. Shen, H. Yun, Z. C. Lipton, Y. Kronrod, and A. Anandkumar, “Deep active learning for named entity recognition,” arXiv preprint arXiv:1707.05928, 2017

    2017 arXiv

  64. [73]

    Latent structured active learning,

    W. Luo, A. Schwing, and R. Urtasun, “Latent structured active learning,” in Int. Conf. Neural Inf. Process. Syst. , vol. 26, 2013

    2013

  65. [74]

    Margin based active learning,

    M.-F. Balcan, A. Broder, and T. Zhang, “Margin based active learning,” in International Conference on Computational Learning Theory . Springer, 2007, pp. 35–50

    2007

  66. [75]

    Panoptic scene graph generation with semantics-prototype learning,

    L. Li, W. Ji, Y. Wu, M. Li, Y. Qin, L. Wei, and R. Zimmermann, “Panoptic scene graph generation with semantics-prototype learning,” in Proc. AAAI Conf. Artif. Intell., vol. 38, no. 4, 2024, pp. 3145–3153

    2024

  67. [76]

    Structured sparse r-cnn for direct scene graph generation,

    Y. Teng and L. Wang, “Structured sparse r-cnn for direct scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2022, pp. 19 437–19 446

    2022

  68. [77]

    Egtr: Extracting graph from transformer for scene graph generation,

    J. Im, J. Nam, N. Park, H. Lee, and S. Park, “Egtr: Extracting graph from transformer for scene graph generation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2024, pp. 24 229–24 238

    2024

  69. [78]

    Adaptive self- training framework for fine-grained scene graph generation,

    K. Kim, K. Yoon, Y. In, J. Moon, D. Kim, and C. Park, “Adaptive self- training framework for fine-grained scene graph generation,” in Proc. Int. Conf. Learn. Represent., 2020, pp. 1–25

    2020

  70. [79]

    Long-tail learning via logit adjustment,

    A. K. Menon, S. Jayasumana, A. S. Rawat, H. Jain, A. Veit, and S. Kumar, “Long-tail learning via logit adjustment,” in Proc. Int. Conf. Learn. Represent., 2020, pp. 1–27. IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 19 APPENDIX FILE TO ‘A R EVERSE CAUSAL FRA...

    2020

  71. [80]

    Enhanced tail relationships. By intervening in the relationship feature space, our method enhances the impact of tail categories on model training, thereby preventing the model from predominantly converging to head categories. Given this, we are not surprised by a decline in t...

  72. [81]

    Some incorrect predictions arise from semantic confusion

    Semantic confusion. Some incorrect predictions arise from semantic confusion. For example, predicting <man, wears, shirt> as <man, wearing, shirt >. While this prediction is §. Ideal training condition refers to the optimizer used for training that can guide the model in reach...

  73. [82]

    Our method demonstrates a strong ability to address noisy labels by predicting fine-grained tail relationships with higher information content

    Annotation granularity. Our method demonstrates a strong ability to address noisy labels by predicting fine-grained tail relationships with higher information content. For instance, it tends to predict <man, sitting on, bench> instead of the coarser <man, on, bench>. This beha...

Pith tools

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