pith. sign in

arxiv: 2606.06369 · v1 · pith:NS5ZPAF5new · submitted 2026-06-04 · 💻 cs.CV

Visual Commonsense Driven Knowledge Refinements for Scene Graph Generation

Ma\"elic Neau , Salim Baloch , Jakob Suchan , Zoe Falomir , Mehul Bhatt This is my paper

Pith reviewed 2026-06-28 01:49 UTC · model grok-4.3

classification 💻 cs.CV
keywords scene graph generationvisual commonsenseknowledge refinementinference-time correctionmodel-agnostic frameworkrelational constraintscommonsense reasoningannotation sparsity
0
0 comments X

The pith

A framework mines visual commonsense constraints from data to refine scene graph predictions at inference time.

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

The paper shows that commonsense constraints capturing spatial, functional, and qualitative relations can be automatically extracted from training data and applied to correct ranked predictions from scene graph generation models. This is done through declarative reasoning at inference without any manual rules or model changes. A reader would care because learning-based models degrade on rare relations due to sparse annotations, and this adds structured knowledge as a complement that works across different models and datasets. The result is consistent gains on standard benchmarks while remaining model-agnostic.

Core claim

We propose a model-agnostic, semantically-guided knowledge refinement framework that systematically mines commonsense-grounded constraints from training data - capturing spatial, functional, and qualitative relational regularities - and uses general declarative commonsense reasoning to correct and refine ranked SGG predictions at inference time. The framework requires no manual rule authoring, no model retraining, and transfers across datasets and architectures. On three standard benchmarks, we obtain consistent improvements over strong baselines, demonstrating that structured visual commonsense reasoning over deep scene semantics is a practical and effective complement to purely learning-ba

What carries the argument

The model-agnostic knowledge refinement framework that mines commonsense-grounded constraints from training data and applies declarative commonsense reasoning to refine ranked predictions at inference.

If this is right

  • Scene graph predictions improve consistently across three standard benchmarks without retraining.
  • The same mined constraints work on different scene graph models and datasets.
  • Annotation sparsity effects are mitigated through post-hoc application of relational constraints.
  • No manual authoring of rules is needed since constraints are derived automatically from data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The refinement step could be inserted into existing pipelines as a lightweight post-processor.
  • Mining constraints from training data alone may limit coverage of rare but valid relations not seen in the data.
  • The approach points toward hybrid systems where learned predictions are adjusted by explicit relational knowledge.
  • Similar constraint mining might apply to other relational vision tasks such as action recognition or visual question answering.

Load-bearing premise

Automatically mined commonsense constraints from the training data capture reliable relational regularities that remain valid and beneficial when applied to predictions from models trained on different data or architectures.

What would settle it

Applying the mined constraints to refine predictions from a new scene graph model on a different benchmark dataset and measuring no gain or a loss in standard metrics such as mean recall would falsify the transfer and improvement claims.

Figures

Figures reproduced from arXiv: 2606.06369 by Jakob Suchan, Ma\"elic Neau, Mehul Bhatt, Salim Baloch, Zoe Falomir.

Figure 1
Figure 1. Figure 1: Sample Refinement Effects for a PSG image. The baseline model (a) predicts three relations that violate commonsense constraints, either because of implausible func￾tional (red) or spatial (blue) constraints. While our refinement (b) eliminates these violations and recovers two missing relations (green edges). applications rely extensively on Scene Graph Generation (SGG) [28] methods to automati￾cally predi… view at source ↗
Figure 2
Figure 2. Figure 2: Visual Commonsense Driven SGG. Conceptual overview of the proposed knowl￾edge refinement method. dataset statistics, or hand-crafted and domain-specific, with no path to open-domain gener￾alisation. Zareian et al. [33] instead learn visual commonsense directly from annotated data, akin to our own data-driven approach, but absorb it into model weights, leaving constraints unverifiable and logically inconsis… view at source ↗
Figure 3
Figure 3. Figure 3: Rule Firing on PSG. Left: firing frequency of each rule type, expressed as the per￾centage of changed pairs in which a rule of that type fires (types can co-fire, so percentages sum to more than 100). Right: distribution of the number of rule types co-firing per changed pair; roughly half of all changed pairs trigger three or more rule types simultaneously [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: τ Tuning. Effect of varying the confidence threshold τ on PSG performance metrics (REACT++ model, PSG dataset). rules, which can be beneficial to account for noise in the data and to avoid overfitting to the training set. We can observe that there is a sweet spot around τ = 0.09 where the F1@K metric is maximized, which indicates that allowing a small amount of violation can lead to better generalization o… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Functional and Spatial Rules - PSG dataset, REACT++ model. 4.4 Qualitative Analyses of Results In this section, we present empirical evidence of our approach, as well as illustration of limitations. In [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: PSG test image 2316458: a bottle and a cake in partial overlap. The two bounding boxes share interior pixels but neither contains the other, yielding RCC5 topology po. B Mined Rule Catalogue From the training annotations we mine three families of commonsense rules: spatial regular￾ities (RCC5 topology, cardinal direction, and bounding-box features), functional constraints, and relational properties. Mining… view at source ↗
Figure 7
Figure 7. Figure 7: 2D Interpretations of the RCC5 Topological Relations. [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The eight cardinal directions used for geometric filtering. The central tile (object [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Rule firing on PSG. Left: firing frequency of each rule type, expressed as the per￾centage of changed pairs in which a rule of that type fires (types can co-fire, so percentages sum to more than 100). Right: distribution of the number of rule types co-firing per changed pair; roughly half of all changed pairs trigger three or more rule types simultaneously. B.5 Relational properties Three syntactic relatio… view at source ↗
Figure 10
Figure 10. Figure 10: The 25 most-frequently firing specific rules on PSG. Rules involving [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Ground-truth outcome by rule type on PSG. Top: full distribution on a log scale; [PITH_FULL_IMAGE:figures/full_fig_p029_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The single rules most responsible for GT-impactful changes on PSG. Left: rules [PITH_FULL_IMAGE:figures/full_fig_p030_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: F1@K contribution per rule type on PSG. riding jumping over about to hit leaning on driving in on back of carrying enclosing touching over swinging sitting on lying on pulling crossing playing with talking to guiding driving on jumping from attached to beside running on slicing feeding parked on looking at drinking hanging from wearing playing painted on in front of holding standing on on flying over thro… view at source ↗
Figure 14
Figure 14. Figure 14: Predicate confusion matrix (original → filtered) for the top-15 most-changed source predicates on PSG. Cell values are the number of pairs whose predicate was reas￾signed from row to column; the colour scale is logarithmic [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative examples of commonsense refinement on PSG. For each image: [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Geometric statistics of subject–object pairs for the PSG dataset. [PITH_FULL_IMAGE:figures/full_fig_p032_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Per-predicate distribution over the five RCC5 topologies for PSG. [PITH_FULL_IMAGE:figures/full_fig_p032_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: PSG predicates associated with the RCC5 discrete (dr) topology – the principal no-overlap case [PITH_FULL_IMAGE:figures/full_fig_p033_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: PSG predicates associated with the RCC5 proper part (pp) topology – the prin￾cipal containment case. −20 −10 0 10 20 30 ΔmR@K (pp) slicing (n=13) playing (n=139) swinging (n=48) hanging from (n=157) holding (n=565) attached to (n=1081) standing on (n=695) enclosing (n=190) beside (n=1778) biting (n=13) looking at (n=467) flying over (n=70) feeding (n=11) falling off (n=3) RCC5 −4 −3 −2 −1 0 1 2 3 ΔmR@K (p… view at source ↗
Figure 20
Figure 20. Figure 20: Per-predicate gains and losses by rule type on PSG, measured as the change in [PITH_FULL_IMAGE:figures/full_fig_p034_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: F1@K contribution per rule type on PSG, grouped by predicate rarity bucket [PITH_FULL_IMAGE:figures/full_fig_p035_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Coverage of predicate-level versus triplet-level rules per rule type on PSG, in [PITH_FULL_IMAGE:figures/full_fig_p035_22.png] view at source ↗
read the original abstract

Learning-driven Scene Graph Generation (SGG) models excel on frequent relation types but degrade sharply under annotation sparsity, failing to capture reliable visual commonsense knowledge. We propose a model-agnostic, semantically-guided knowledge refinement framework that systematically mines commonsense-grounded constraints from training data - capturing spatial, functional, and qualitative relational regularities - and uses general declarative commonsense reasoning to correct and refine ranked SGG predictions at inference time. The framework requires no manual rule authoring, no model retraining, and transfers across datasets and architectures. On three standard benchmarks, we obtain consistent improvements over strong baselines, demonstrating that structured visual commonsense reasoning over deep scene semantics is a practical and effective complement to purely learning-based scene graph generation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces a model-agnostic, semantically-guided knowledge refinement framework for Scene Graph Generation (SGG). It automatically mines commonsense-grounded constraints from training data that capture spatial, functional, and qualitative relational regularities, then applies general declarative commonsense reasoning to correct and refine ranked SGG predictions at inference time. The framework requires no manual rule authoring or model retraining and is claimed to transfer across datasets and architectures, yielding consistent improvements over strong baselines on three standard benchmarks.

Significance. If the transfer results hold under rigorous cross-distribution testing, the work would demonstrate a practical, training-free mechanism for injecting structured visual commonsense into learning-based SGG pipelines. This hybrid approach could meaningfully address annotation sparsity without retraining costs, offering a reusable complement to purely data-driven methods in relational scene understanding tasks.

major comments (1)
  1. [Abstract and experiments section] Abstract and experiments section: the central claim that automatically mined constraints encode transferable visual commonsense (rather than source-dataset annotation biases) is load-bearing for the transfer assertion. The manuscript must report explicit cross-dataset and cross-architecture quantitative results showing that constraints mined from one training distribution improve predictions from models trained on different distributions; without such controls, the reported gains may not generalize beyond the source statistics.
minor comments (1)
  1. Clarify in the method description whether the declarative reasoning step uses any dataset-specific thresholds or post-processing that could affect reproducibility across new architectures.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the importance of rigorous controls to support the transferability claim. We agree that explicit cross-dataset and cross-architecture experiments are required to distinguish visual commonsense from source-specific annotation biases, and we will incorporate them in the revision.

read point-by-point responses

  1. Referee: [Abstract and experiments section] Abstract and experiments section: the central claim that automatically mined constraints encode transferable visual commonsense (rather than source-dataset annotation biases) is load-bearing for the transfer assertion. The manuscript must report explicit cross-dataset and cross-architecture quantitative results showing that constraints mined from one training distribution improve predictions from models trained on different distributions; without such controls, the reported gains may not generalize beyond the source statistics.

    Authors: We concur that the current experiments, while showing gains across three benchmarks, do not include the precise controls requested (constraints mined from distribution A applied to a model trained on distribution B). In the revised version we will add: (1) cross-dataset tests mining constraints from Visual Genome and applying them to models trained on GQA and Open Images; (2) cross-architecture tests using the same mined constraints on multiple SGG backbones (e.g., Motifs, VCTree, and a transformer-based model). These results will be reported in a new subsection of the experiments with quantitative tables and analysis of whether improvements persist under distribution shift. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation is self-contained empirical claim

full rationale

The provided abstract and description contain no equations, fitting procedures, or self-citations that reduce any claimed result to its inputs by construction. The framework is presented as mining constraints from training data and applying them at inference time to obtain empirical improvements on benchmarks, with the transfer property asserted as an observed outcome rather than a definitional or fitted necessity. No load-bearing step equates a prediction to a parameter fit or renames an input as output. This is the expected non-finding for a high-level methods description lacking mathematical reductions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents enumeration of free parameters, axioms, or invented entities; none are explicitly stated in the provided text.

pith-pipeline@v0.9.1-grok · 5658 in / 1043 out tokens · 24190 ms · 2026-06-28T01:49:14.869520+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 10 canonical work pages

  1. [1]

    Og-sgg: ontology-guided scene graph generation—a case study in transfer learning for telepresence robotics.IEEE Access, 10:132564–132583, 2022

    Fernando Amodeo, Fernando Caballero, Natalia Díaz-Rodríguez, and Luis Merino. Og-sgg: ontology-guided scene graph generation—a case study in transfer learning for telepresence robotics.IEEE Access, 10:132564–132583, 2022

    2022

  2. [2]

    3d scene graph: A structure for unified semantics, 3d space, and camera

    Iro Armeni, Zhi-Yang He, JunYoung Gwak, Amir R Zamir, Martin Fischer, Jitendra Malik, and Silvio Savarese. 3d scene graph: A structure for unified semantics, 3d space, and camera. InProceedings of the IEEE/CVF international conference on computer vision, pages 5664–5673, 2019

    2019

  3. [3]

    Artificial visual intelligence - perceptual common- sense for human-centred cognitive technologies

    Mehul Bhatt and Jakob Suchan. Artificial visual intelligence - perceptual common- sense for human-centred cognitive technologies. In Mohamed Chetouani, Virginia Dignum, Paul Lukowicz, and Carles Sierra, editors,Human-Centered Artificial In- telligence - Advanced Lectures, 18th European Advanced Course on AI, ACAI 2021, Berlin, Germany, October 11-15, 2021...

    work page doi:10.1007/978-3-031-24349-3 2021

  4. [4]

    Cohn, Brandon Bennett, John Gooday, and Nicholas M

    Anthony G. Cohn, Brandon Bennett, John Gooday, and Nicholas M. Gotts. Qualitative spatial representation and reasoning with the region connection calculus.GeoInformat- ica, 1(3):275–316, 1997

    1997

  5. [5]

    Commonsense reasoning and commonsense knowledge in artificial intelligence.Communications of the ACM, 58(9):92–103, 2015

    Ernest Davis and Gary Marcus. Commonsense reasoning and commonsense knowledge in artificial intelligence.Communications of the ACM, 58(9):92–103, 2015

    2015

  6. [6]

    Explainable zero-shot visual question answering via logic-based reasoning

    Thomas Eiter, Jan Hadl, Nelson Higuera Ruiz, Lukas Lange, Johannes Oetsch, Bileam Scheuvens, and Jannik Strötgen. Explainable zero-shot visual question answering via logic-based reasoning. In Leilani H. Gilpin, Eleonora Giunchiglia, Pascal Hitzler, and Emile van Krieken, editors,Proceedings of The 19th International Conference on Neu- rosymbolic Learning ...

    2025

  7. [7]

    Clingo = ASP + control: Preliminary report.CoRR, abs/1405.3694, 2014

    Martin Gebser, Roland Kaminski, Benjamin Kaufmann, and Torsten Schaub. Clingo = ASP + control: Preliminary report.CoRR, abs/1405.3694, 2014

    Pith/arXiv arXiv 2014

  8. [8]

    Theory Solving Made Easy with Clingo 5

    Martin Gebser, Roland Kaminski, Benjamin Kaufmann, Max Ostrowski, Torsten Schaub, and Philipp Wanko. Theory Solving Made Easy with Clingo 5. In Manuel Carro, Andy King, Neda Saeedloei, and Marina De V os, editors,Technical Communi- cations of the 32nd International Conference on Logic Programming (ICLP 2016), vol- ume 52 ofOpenAccess Series in Informatics...

    work page doi:10.4230/oasics.iclp.2016.2 2016

  9. [9]

    Goyal and Max J

    Roop K. Goyal and Max J. Egenhofer. Similarity of cardinal directions. In Christian S. Jensen, Markus Schneider, Bernhard Seeger, and Vassilis J. Tsotras, editors,Advances in Spatial and Temporal Databases, pages 36–55, Berlin, Heidelberg, 2001. Springer Berlin Heidelberg. ISBN 978-3-540-47724-2

    2001

  10. [10]

    Enhancing scene graph generation with hierarchical relationships and commonsense knowledge

    Bowen Jiang, Zhijun Zhuang, Shreyas S Shivakumar, and Camillo J Taylor. Enhancing scene graph generation with hierarchical relationships and commonsense knowledge. NEAU ET. AL.: COMMONSENSE DRIVEN KNOWLEDGE REFINEMENTS FOR SGG17 In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 8883–8894. IEEE, 2025

    2025

  11. [11]

    Image generation from scene graphs

    Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018

    2018

  12. [12]

    A survey of neu- rosymbolic visual reasoning with scene graphs and common sense knowledge.Neu- rosymbolic Artificial Intelligence, 1:NAI–240719, 2025

    M Jaleed Khan, Filip Ilievski, John G Breslin, and Edward Curry. A survey of neu- rosymbolic visual reasoning with scene graphs and common sense knowledge.Neu- rosymbolic Artificial Intelligence, 1:NAI–240719, 2025

    2025

  13. [13]

    Expressive scene graph generation using commonsense knowledge infusion for visual understanding and rea- soning

    Muhammad Jaleed Khan, John G Breslin, and Edward Curry. Expressive scene graph generation using commonsense knowledge infusion for visual understanding and rea- soning. InEuropean Semantic Web Conference, pages 93–112. Springer, 2022

    2022

  14. [14]

    Breslin, and Edward Curry

    Muhammad Jaleed Khan, John G. Breslin, and Edward Curry. Knowzrel: Common sense knowledge-based zero-shot relationship retrieval for generalized scene graph generation.IEEE Transactions on Artificial Intelligence, 6(12):3184–3194, 2025. doi: 10.1109/TAI.2025.3544177

    work page doi:10.1109/tai.2025.3544177 2025

  15. [15]

    Visual genome: Connecting language and vision using crowdsourced dense image annotations.Inter- national journal of computer vision, 123(1):32–73, 2017

    Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations.Inter- national journal of computer vision, 123(1):32–73, 2017

    2017

  16. [16]

    Visual question an- swering over scene graph

    Soohyeong Lee, Ju-Whan Kim, Youngmin Oh, and Joo Hyuk Jeon. Visual question an- swering over scene graph. In2019 First International Conference on Graph Computing (GC), pages 45–50. IEEE, 2019

    2019

  17. [17]

    Springer, 2019

    Vladimir Lifschitz.Answer Set Programming. Springer, 2019. ISBN 978-3-030- 24657-0. doi: 10.1007/978-3-030-24658-7

    work page doi:10.1007/978-3-030-24658-7 2019

  18. [18]

    Wiley, 2013

    Gerard Ligozat.Qualitative Spatial and Temporal Reasoning. Wiley, 2013. ISBN 9781118601457. doi: https://doi.org/10.1002/9781118601457.fmatter

    work page doi:10.1002/9781118601457.fmatter 2013

  19. [19]

    React++: Efficient cross-attention for real-time scene graph generation.arXiv preprint arXiv:2603.06386, 2026

    Maëlic Neau and Zoe Falomir. React++: Efficient cross-attention for real-time scene graph generation.arXiv preprint arXiv:2603.06386, 2026

    arXiv 2026

  20. [20]

    In defense of scene graph generation for human-robot open-ended interaction in service robotics

    Maëlic Neau, Paulo Santos, Anne-Gwenn Bosser, and Cédric Buche. In defense of scene graph generation for human-robot open-ended interaction in service robotics. In Robot World Cup, pages 299–310. Springer, 2023

    2023

  21. [21]

    React: Real-time efficiency and accuracy compromise for tradeoffs in scene graph generation

    Maëlic Neau, Paulo Eduardo Santos, Anne-Gwenn Bosser, Akihiro Sugimoto, and Cedric Buche. React: Real-time efficiency and accuracy compromise for tradeoffs in scene graph generation. In36th British Machine Vision Conference 2025, BMVC 2025, Sheffield, UK, November 24-27, 2025. BMV A, 2025

    2025

  22. [22]

    Symbolic rule extraction from attention-guided sparse representations in vision transformers.Theory and Practice of Logic Programming, 25 (4):722–738, 2025

    Parth Padalkar and Gopal Gupta. Symbolic rule extraction from attention-guided sparse representations in vision transformers.Theory and Practice of Logic Programming, 25 (4):722–738, 2025. doi: 10.1017/S1471068425100318. 18NEAU ET. AL.: COMMONSENSE DRIVEN KNOWLEDGE REFINEMENTS FOR SGG

    work page doi:10.1017/s1471068425100318 2025

  23. [23]

    Semantic question-answering with video and eye- tracking data: AI foundations for human visual perception driven cognitive film studies

    Jakob Suchan and Mehul Bhatt. Semantic question-answering with video and eye- tracking data: AI foundations for human visual perception driven cognitive film studies. In Subbarao Kambhampati, editor,Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, IJCAI 2016, New York, NY, USA, 9-15 July 2016, pages 2633–2639. IJC...

    2016

  24. [24]

    Commonsense visual sense- making for autonomous driving - on generalised neurosymbolic online abduction inte- grating vision and semantics.Artif

    Jakob Suchan, Mehul Bhatt, and Srikrishna Varadarajan. Commonsense visual sense- making for autonomous driving - on generalised neurosymbolic online abduction inte- grating vision and semantics.Artif. Intell., 299:103522, 2021. doi: 10.1016/J.ARTINT. 2021.103522

    work page doi:10.1016/j.artint 2021

  25. [25]

    ASP-driven visual commonsense: a general framework for reasoning about embodied interaction in the wild

    Jakob Suchan, Mehul Bhatt, and Julius Monsen. ASP-driven visual commonsense: a general framework for reasoning about embodied interaction in the wild. InProceed- ings of the 22nd International Conference on Principles of Knowledge Representation and Reasoning, KR ’25, 2025. ISBN 978-1-956792-08-9. doi: 10.24963/kr.2025/61

    work page doi:10.24963/kr.2025/61 2025

  26. [26]

    Unbi- ased scene graph generation from biased training

    Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbi- ased scene graph generation from biased training. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716–3725, 2020

    2020

  27. [27]

    Yolov12: Attention-centric real-time object detectors.Advances in neural information processing systems, 38:78433–78457, 2026

    Yunjie Tian, Qixiang Ye, and David Doermann. Yolov12: Attention-centric real-time object detectors.Advances in neural information processing systems, 38:78433–78457, 2026

    2026

  28. [28]

    Scene graph generation by iterative message passing

    Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017

    2017

  29. [29]

    Panoptic scene graph generation

    Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. InEuropean conference on computer vision, pages 178–196. Springer, 2022

    2022

  30. [30]

    Auto-encoding scene graphs for image captioning

    Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019

    2019

  31. [31]

    Neurasp: Embracing neural networks into answer set programming

    Zhun Yang, Adam Ishay, and Joohyung Lee. Neurasp: Embracing neural networks into answer set programming. In Christian Bessiere, editor,Proceedings of the Twenty- Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 1755–

  32. [32]

    doi: 10.24963/ijcai.2020/243

    International Joint Conferences on Artificial Intelligence Organization, 7 2020. doi: 10.24963/ijcai.2020/243. Main track

    work page doi:10.24963/ijcai.2020/243 2020

  33. [33]

    Generative visual commonsense reasoning with scene graphs

    Fan Yuan, Xiaoyuan Fang, Rong Quan, Jing Li, Wei Bi, Xiaogang Xu, and Piji Li. Generative visual commonsense reasoning with scene graphs. Preprint, 2025

    2025

  34. [34]

    Learning visual commonsense for robust scene graph generation

    Alireza Zareian, Zhecan Wang, Haoxuan You, and Shih-Fu Chang. Learning visual commonsense for robust scene graph generation. InEuropean Conference on Computer Vision, pages 642–657. Springer, 2020

    2020

  35. [35]

    Neural motifs: Scene graph parsing with global context

    Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5831–5840, 2018. NEAU ET. AL.: COMMONSENSE DRIVEN KNOWLEDGE REFINEMENTS FOR SGG19

    2018

  36. [36]

    From recognition to cog- nition: Visual commonsense reasoning

    Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cog- nition: Visual commonsense reasoning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6720–6731, 2019. doi: 10.1109/CVPR.2019.00688. 20NEAU ET. AL.: COMMONSENSE DRIVEN KNOWLEDGE REFINEMENTS FOR SGG Contents This supplement p...

    work page doi:10.1109/cvpr.2019.00688 2019