Pith. sign in

REVIEW 4 major objections 3 minor 48 references

Taking A Closer Look at Interacting Objects: Interaction-Aware Open Vocabulary Scene Graph Generation

T0 review · 4 major / 3 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read INOVA claims that injecting interaction awareness into pre-training, fine-tuning, and distillation lifts open-vocabulary scene graph generation to state-of-the-art results on VG and GQA.

desk verdict Solid incremental OVSGG paper with credible VG gains, but the missing GQA results and misattributed ablation numbers keep it from being a clean accept. read the letter →

arxiv 2502.03856 v1 pith:H7UMAGKV submitted 2025-02-06 cs.CV

classification cs.CV
keywords openvocabularyscenegraphgenerationinteraction-awaretargetinteraction-guidedqueryselectionrelative-interactionretentiondistillationbipartitematchingweaklysupervisedpre-trainingknowledgevisualgenome
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

Open-vocabulary scene graph generation aims to extract, from an image, a structured graph of objects and their relationships using categories that may never have been seen during training. This paper argues that existing two-stage pipelines—weakly supervised pre-training on captions followed by supervised fine-tuning on annotated scene graphs—fail because they treat every object instance as equivalent, so interacting pairs and idle bystanders compete for the same supervision and matching slots. To fix this, INOVA inserts interaction awareness at each stage: bidirectional interaction prompts ground the right subject–object pairs during pre-training, an interaction-guided query selection prioritizes participants in relations during fine-tuning, and a relative-interaction knowledge distillation keeps interacting pairs distinct from the background. On the VG benchmark, INOVA reports state-of-the-art results, with the largest gains on novel relations, supporting the paper's central claim that explicit interaction modeling is what the task was missing.

What carries the argument

The argument is carried by three mechanisms. Interaction-aware target generation builds bidirectional interaction prompts—'man hold surfboard' and 'surfboard held by man', with the reverse phrase written by an LLM—and feeds them to Grounding DINO so that the text attention enriches each object token with its relational context; overlapping subject and object boxes are then combined by IoU into triplet supervision. Interaction-guided query selection scores every visual token by its maximum similarity to object and relation class tokens, picks the top K to initialize queries, then re-encodes predicted triplets as interaction prompts and re-ranks tokens by interaction relevance, reserving remaining query slots for objects missed by the first pass. Interaction-consistent knowledge distillation adds a relative-interaction retention loss that aligns the L2-normalized pairwise similarity matrices of teacher and student edge features, preserving the structure that separates true interaction pairs from background pairs.

What would settle it

Compute the precision of the rule-based pseudo-triplets on a random sample of VG or COCO images by comparing IoU-combined subject–object boxes against human-annotated triplets; if a substantial share of overlaps pair non-interacting instances (e.g., two people overlapping one surfboard while only one holds it), the pre-training supervision is noisy and part of the reported gain may come from the IoU filter rather than from interaction awareness. Replacing the LLM counter-actions with grammatically valid random reverse phrases and observing no drop in novel-relation recall would likewise refute the claim that semantically accurate bidirectional prompts are what carries the effect.

Watch

Extended reading notes

Core claim

The paper's central claim is that the persistent mismatch between interacting and non-interacting objects—not a lack of model capacity or annotation quality—is a primary cause of errors in open-vocabulary scene graph generation, and that injecting interaction awareness into every training stage removes that obstacle. In the OvD+R-SGG setting on VG150 with a Swin-T backbone, INOVA reaches 21.27 R@100 on the joint base+novel split, 21.10 on novel objects, and 19.46 on novel relations, improving over the OvSGTR baseline by +4.90, +3.62, and +8.28 R@100 respectively; ablations attribute these gains to the three interaction-aware components, with the full combination giving the best numbers. The paper states this as consistent state-of-the-art performance across all reported metrics and both evaluated benchmarks.

Load-bearing premise

The load-bearing premise is that the pseudo-supervision is clean: after grounding the two prompt directions, any subject and object boxes that overlap are treated as a true interacting pair, and the LLM-generated reverse phrases are assumed to be semantically accurate, with no measurement or filtering of the mistakes this can produce.

Editorial extensions

If this is right

  • Pre-training supervision becomes cleaner: grounding with bidirectional prompts localizes the subject and object of a parsed relation instead of all category instances, so the weakly supervised stage produces fewer mismatched triplets.
  • Fine-tuning becomes more selective: query initialization and refinement favor objects that participate in relations, so bipartite graph matching is less likely to pair a non-interacting instance with a ground-truth relation target.
  • Novel relation recall improves most: the largest reported gains are on unseen relations (+8.28 R@100 over OvSGTR in OvD+R-SGG), suggesting the interaction-aware objectives specifically help generalization rather than memorization of base classes.
  • The three components are complementary and additive: ablations show each of ITG, IQS, and RRD contributes alone, and their combination yields the best joint base+novel, novel-object, and novel-relation recall.
  • The pre-trained model transfers directly: INOVA pre-trained on COCO captions with interaction-aware targets achieves higher zero-shot recall on VG150 than the corresponding baseline, indicating the pretraining signal is the foundation for the later gains.

Reading between the lines

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

  • A testable extension is to ablate the LLM-written counter-actions by replacing them with random reverse phrases; if bidirectional prompts matter only through their interaction context, random counter-actions that keep the same grammar should give most of the benefit, whereas if semantic accuracy of the reverse relation is load-bearing, performance should drop.
  • The IoU-based combination of overlapping boxes functions as a denoising filter on pseudo-labels, so part of the ITG gain may reflect cleaner supervision rather than interaction reasoning per se; isolating these two effects would require comparing against a variant that filters by category co-occurrence statistics instead of interaction prompts.
  • The same interaction-aware design could transfer to other end-to-end detection tasks that use bipartite matching, such as open-vocabulary detection or panoptic scene graph generation, wherever the teacher signal is a triplet or pair relation rather than a single object class.
  • A stricter check of the paper's premise would measure, on a sample of images, how often the rule-based pseudo-triplets actually match human-annotated interactions; that number would show how much of the reported improvement is attributable to interaction modeling versus supervision denoising.
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 / 3 minor

Summary. The paper proposes INOVA, an interaction-aware open-vocabulary scene graph generation framework. It introduces three components: interaction-aware target generation (ITG) during weak pre-training, which uses bidirectional interaction prompts and IoU-based combination of detected boxes to form triplet supervision; interaction-guided query selection (IQS) during supervised fine-tuning, which ranks and selects visual tokens by interaction relevance; and relative-interaction retention distillation (RRD), which preserves inter-pair relational structure from a self-distillation teacher. The authors report state-of-the-art results on VG150 in both OvR-SGG and OvD+R-SGG settings, with the largest gains on novel relation recall. The abstract claims experiments on VG and GQA, but the provided manuscript contains VG results only.

Significance. If the results hold, INOVA addresses a genuine limitation of existing OVSGG pipelines—the uniform treatment of interacting and non-interacting objects—and demonstrates that explicit interaction modeling can yield substantial gains, especially for novel relation prediction. The bidirectional prompting and interaction-consistent distillation ideas are technically plausible and could inspire further work. The VG results are consistently reported and the ablation structure is informative. However, the significance is currently undercut by three issues: the GQA evaluation is missing from the manuscript, the component-wise attribution in Section 4.3 is contradicted by Table 3, and the core ITG supervision assumption (IoU overlap implies interaction) is not validated. These issues are fixable, and the underlying method appears sound enough to merit a major revision rather than rejection.

major comments (4)
  1. [Sec. 4.3, Table 3] The text attributes a 3.94% R@100 gain to ITG and a 2.83% gain to RRD, but Table 3 shows the opposite: the ITG-only row (19.20) differs from the baseline (16.37) by 2.83, and the RRD-only row (20.31) differs by 3.94. This misattribution is load-bearing because the central claim is that each interaction-aware component causes the observed improvements. Please correct the text or the table and re-verify the component-wise conclusions.
  2. [Abstract, Sec. 4.1] The abstract and Section 4.1 claim evaluation on GQA, but no GQA results, tables, or setting details appear in the manuscript. The only mention is a brief reference to Appendix A. Since the abstract states that INOVA achieves state-of-the-art performance on two benchmarks, the missing GQA results make this claim unverifiable. Please include the GQA experiments in the main text or clearly point to a non-empty appendix and summarize the key numbers.
  3. [Sec. 3.2.1, Table 3] ITG builds triplet pseudo-labels by combining subject and object boxes whose IoU exceeds an unspecified threshold, effectively assuming that overlapping boxes indicate a true interaction. The paper does not report the IoU threshold, the precision of generated triplets, or any sensitivity analysis. Given that Grounding DINO can produce multiple or spurious boxes (as Figure 2 concedes), the 2.83-point R@100 gain attributed to ITG could stem from filtering low-overlap pairs rather than from the bidirectional interaction prompts encoding relational semantics. Please validate the interaction assumption, e.g., by reporting pseudo-label precision, varying the IoU threshold, and comparing against a non-interaction-aware filtering baseline.
  4. [Sec. 3.2, Sec. 4.1] The free hyperparameters γ (Eq. 1), K (Eq. 2), L (Eqs. 4-6), β1, and β2 (Eq. 10) are never assigned values in the main text, and the appendix that supposedly contains implementation details is not included in the manuscript. Without these values, the experiments are not reproducible. Please report all hyperparameter settings, ideally in the main text or in an included appendix.
minor comments (3)
  1. [Throughout] The paper has several typographical and formatting issues: 'INOV A' instead of 'INOVA' in the body, a missing space in '1The Hong Kong...' on page 1, and the open 'man ' artifacts in Figure 1. These should be cleaned up.
  2. [Sec. 4.1] The metric definitions say 'Mean R@K (mR@K): The average R@K across all categories,' but the paper never reports mR@K in any table. Please either add these numbers or remove the mR@K definition.
  3. [Sec. 4.3] The sentence in the Diagnostic Experiment says 'the improvement is less pronounced than expected, since each strategy prioritizes interacting objects, which may lead to diminishing returns.' This is speculative and not backed by any measurement of diminishing returns; please rephrase or support with data.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the SOTA claim is supported by external benchmark evaluation and standard ablations, with no derivation reducing to its inputs.

full rationale

This is an empirical systems paper rather than a formal derivation chain. The central claim, that INOVA achieves state-of-the-art open-vocabulary scene graph generation, is supported by direct measurements on fixed external benchmarks (VG and GQA) against published baselines, so the result is not defined in terms of its own inputs. The three proposed components (ITG, IQS, RRD) are described operationally and their contributions are tested by ablations on the same test set; this is the standard experimental method of causal attribution, not a fitted parameter renamed as a prediction. The interaction-aware target generation uses rule-based IoU combination of grounded boxes to build pre-training supervision, but the final evaluation uses ground-truth triplet annotations, so the training signal is not identical to the test target. Self-references to prior work by the same group (Li et al. 2022b, Li et al. 2024a) appear only as related-work context or design inspiration and are not load-bearing for the reported improvement. The KD teacher is indeed a pretrained version of the same model, but self-distillation is used as a regularizer and the gain is evaluated externally, so it does not reduce to the teacher's output by construction. One numerical inconsistency exists in Section 4.3: the text attributes a 3.94% R@100 gain to ITG, whereas Table 3 shows the ITG-only row gains 2.83% and the RRD-only row gains 3.94%. This is an accuracy or reporting issue in the ablation narrative, not a circular step. No uniqueness theorems, forced definitions, or fitted-to-output predictions are present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

All free parameters are hyperparameters with no reported values; the central claim depends on them but they are deferred to the missing appendix. No new physical entities are introduced. The dominant assumptions are domain modeling choices inherited from weakly supervised SGG and a new IoU-overlap proxy for interactions.

free parameters (5)
  • gamma (Eq. 1)
    Balances object-token vs relation-token similarity in the relevance score; value not reported or ablated.
  • K (Eq. 2, 6)
    Total number of visual queries selected for decoding; value not reported.
  • L (Eq. 4-6)
    Number of interaction-relevant queries selected in Step II; value not reported.
  • beta1 (Eq. 10)
    Weight for visual-concept retention distillation loss; value not reported.
  • beta2 (Eq. 10)
    Weight for relative-interaction retention distillation loss; value not reported.
assumptions (4)
  • domain assumption Pre-trained backbones (Swin, BERT, Grounding DINO) provide reliable visual and text representations for grounding.
    The method builds on these models without re-training them; if their representations fail on novel categories, the interaction-aware modules inherit the failure.
  • domain assumption Caption graphs parsed by the Schuster parser and grounded by the detector are sufficient supervision for pre-training.
    Sec. 3.2.1 follows prior weakly supervised SGG; the paper does not validate the parser or detector noise on the final task.
  • ad hoc to paper IoU overlap between subject and object boxes indicates a true interaction.
    Sec. 3.2.1 rule-based combination; no validation of this proxy, which is a new assumption introduced for ITG.
  • domain assumption Teacher model edge features on negative samples are a reliable relational prior for distillation.
    Sec. 3.2.3, VRD and RRD assume the pre-trained teacher's edge space is trustworthy and useful for preserving relational structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Taking A Closer Look at Interacting Objects: Interaction-Aware Open Vocabulary Scene Graph Generation." pith.science (2026). https://pith.science/paper/H7UMAGKV

@misc{pith2026250203856,
  author       = {Pith},
  title        = {Pith review of: Taking A Closer Look at Interacting Objects: Interaction-Aware Open Vocabulary Scene Graph Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H7UMAGKV}},
  note         = {Machine review of arXiv:2502.03856}
}
read the original abstract

Today's open vocabulary scene graph generation (OVSGG) extends traditional SGG by recognizing novel objects and relationships beyond predefined categories, leveraging the knowledge from pre-trained large-scale models. Most existing methods adopt a two-stage pipeline: weakly supervised pre-training with image captions and supervised fine-tuning (SFT) on fully annotated scene graphs. Nonetheless, they omit explicit modeling of interacting objects and treat all objects equally, resulting in mismatched relation pairs. To this end, we propose an interaction-aware OVSGG framework INOVA. During pre-training, INOVA employs an interaction-aware target generation strategy to distinguish interacting objects from non-interacting ones. In SFT, INOVA devises an interaction-guided query selection tactic to prioritize interacting objects during bipartite graph matching. Besides, INOVA is equipped with an interaction-consistent knowledge distillation to enhance the robustness by pushing interacting object pairs away from the background. Extensive experiments on two benchmarks (VG and GQA) show that INOVA achieves state-of-the-art performance, demonstrating the potential of interaction-aware mechanisms for real-world applications.

Figures

Figures reproduced from arXiv: 2502.03856 by the authors.

Figure 1
Figure 1. Overview of the OVSGG framework challenges. 1) VLM Pre-training, using solely entity categories for object detec￾tion causes ambiguity in associating object pairs (e.g., identifying the correct “man-surfboard” for the “hold”). 2) SFT, bipartite graph matching misaligns non-interacting objects (e.g., “man ”) with interacting target “man” in ⟨man, riding, horse⟩. It pushes beyond predefined categories to support the r… view at source ↗
Figure 2
Figure 2. Overview of INOVA for OVSGG. (a) VLM Pre-training: Interaction-aware target generation uses bidirectional interaction prompts and rule-based bounding box combinations to generate supervision, enriching object tokens with contextual interaction semantics. (b) SFT: A two-step interaction-guided query selection (IQS) prioritizes interacting objects and integrates relational context into object tokens, refining queries … view at source ↗
Figure 3
Figure 3. Illustration of interaction-consistent KD. interacting pairs from background pairs and address catas￾trophic forgetting of learned relational semantics mentioned in (Chen et al., 2024b). Specifically, it leverages the VLM pre-trained in the first stage as the teacher model. The stu￾dent network is designed as a pseudo-siamese structure of the teacher model, initialized with the teacher’s parameters. Interaction-cons… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Interaction-aware target generation. significant performance boosts. RRD contributes 2.83% R@100 gains, improving the model’s ability to handle novel classes effectively. Fourth, the integration of all three com￾ponents (i.e., ITG, IQS, and RRD) yields the best overall…
Figure 5
Figure 5. Figure 5: Interaction-guided query selection. variants consistently outperform others, achieving 13.31% R@100 with Swin-T and 14.22% R@100 with Swin-B. These results demonstrate the effectiveness of incorporating ITG in the VLM pre-training process. In addition, we visualized th…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 38 canonical work pages

  1. [1]

    U., Khan, S

    Bangalath, H., Maaz, M., Khattak, M. U., Khan, S. H., and Shahbaz Khan, F. Bridging the gap between object and image-level representations for open-vocabulary detection. In NeurIPS, volume 35, pp.\ 33781--33794, 2022

  2. [2]

    End-to-end object detection with transformers

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., and Zagoruyko, S. End-to-end object detection with transformers. In ECCV, pp.\ 213--229, 2020

  3. [3]

    Scene graph generation with role-playing large language models

    Chen, G., Li, J., and Wang, W. Scene graph generation with role-playing large language models. NeurIPS, 2024 a

  4. [4]

    Chen, X., Fang, H., Lin, T., Vedantam, R., Gupta, S., Doll \' a r, P., and Zitnick, C. L. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015

  5. [5]

    E., Ahmed, F., Gan, Z., Cheng, Y., and Liu, J

    Chen, Y., Li, L., Yu, L., Kholy, A. E., Ahmed, F., Gan, Z., Cheng, Y., and Liu, J. UNITER: universal image-text representation learning. In ECCV, pp.\ 104--120, 2020

  6. [6]

    Gpt4sgg: Synthesizing scene graphs from holistic and region-specific narratives

    Chen, Z., Wu, J., Lei, Z., Zhang, Z., and Chen, C. Gpt4sgg: Synthesizing scene graphs from holistic and region-specific narratives. arXiv preprint arXiv:2312.04314, 2023

  7. [7]

    Expanding scene graph boundaries: Fully open-vocabulary scene graph generation via visual-concept alignment and retention

    Chen, Z., Wu, J., Lei, Z., Zhang, Z., and Chen, C. Expanding scene graph boundaries: Fully open-vocabulary scene graph generation via visual-concept alignment and retention. In ECCV, 2024 b

  8. [8]

    BERT: pre-training of deep bidirectional transformers for language understanding

    Devlin, J., Chang, M., Lee, K., and Toutanova, K. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, pp.\ 4171--4186, 2019

Show all 48 references
  1. [9]

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

    Dong, X., Gan, T., Song, X., Wu, J., Cheng, Y., and Nie, L. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In CVPR, pp.\ 19427--19436, 2022

  2. [10]

    Open-vocabulary object detection via vision and language knowledge distillation

    Gu, X., Lin, T.-Y., Kuo, W., and Cui, Y. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021

  3. [11]

    Towards open-vocabulary scene graph generation with prompt-based finetuning

    He, T., Gao, L., Song, J., and Li, Y. Towards open-vocabulary scene graph generation with prompt-based finetuning. In ECCV, pp.\ 56--73, 2022

  4. [12]

    Hudson, D. A. and Manning, C. D. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pp.\ 6700--6709, 2019

  5. [13]

    Llm4sgg: Large language models for weakly supervised scene graph generation

    Kim, K., Yoon, K., Jeon, J., In, Y., Moon, J., Kim, D., and Park, C. Llm4sgg: Large language models for weakly supervised scene graph generation. In CVPR, pp.\ 28306--28316, 2024

  6. [14]

    A., et al

    Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.-J., Shamma, D. A., et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123: 0 32--73, 2017

  7. [15]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Li, J., Li, D., Savarese, S., and Hoi, S. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pp.\ 19730--19742. PMLR, 2023

  8. [16]

    Zero-shot visual relation detection via composite visual cues from large language models

    Li, L., Xiao, J., Chen, G., Shao, J., Zhuang, Y., and Chen, L. Zero-shot visual relation detection via composite visual cues from large language models. NeurIPS, 36, 2024 a

  9. [17]

    H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J., Chang, K., and Gao, J

    Li, L. H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J., Chang, K., and Gao, J. Grounded language-image pre-training. In CVPR, pp.\ 10955--10965, 2022 a

  10. [18]

    From pixels to graphs: Open-vocabulary scene graph generation with vision-language models

    Li, R., Zhang, S., Lin, D., Chen, K., and He, X. From pixels to graphs: Open-vocabulary scene graph generation with vision-language models. In CVPR, pp.\ 28076--28086, 2024 b

  11. [19]

    Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation

    Li, X., Chen, L., Ma, W., Yang, Y., and Xiao, J. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. In ACMMM, pp.\ 4204--4213, 2022 b

  12. [20]

    B., He, K., and Doll \' a r, P

    Lin, T., Goyal, P., Girshick, R. B., He, K., and Doll \' a r, P. Focal loss for dense object detection. In ICCV, pp.\ 2999--3007, 2017

  13. [21]

    Liu, H., Li, C., Wu, Q., and Lee, Y. J. Visual instruction tuning. NeurIPS, 36, 2024

  14. [22]

    Grounding DINO: marrying DINO with grounded pre-training for open-set object detection

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J., and Zhang, L. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. CoRR, abs/2303.05499, 2023

  15. [23]

    Relation-aware hierarchical prompt for open-vocabulary scene graph generation

    Liu, T., Li, R., Wang, C., and He, X. Relation-aware hierarchical prompt for open-vocabulary scene graph generation. In AAAI, 2025

  16. [24]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., and Guo, B. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pp.\ 9992--10002, 2021

  17. [25]

    Gpt-4 technical report

    OpenAI, R. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2 0 (5), 2023

  18. [26]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In ICML, pp.\ 8748--8763, 2021

  19. [27]

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

    Ren, S., He, K., Girshick, R., and Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015

  20. [28]

    D., and Savarese, S

    Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I. D., and Savarese, S. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pp.\ 658--666, 2019

  21. [29]

    Schuster, S., Krishna, R., Chang, A., Fei-Fei, L., and Manning, C. D. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pp.\ 70--80, 2015

  22. [30]

    A simple baseline for weakly-supervised scene graph generation

    Shi, J., Zhong, Y., Xu, N., Li, Y., and Xu, C. A simple baseline for weakly-supervised scene graph generation. In ICCV, pp.\ 16393--16402, 2021

  23. [31]

    Relationformer: A unified framework for image-to-graph generation

    Shit, S., Koner, R., Wittmann, B., Paetzold, J., Ezhov, I., Li, H., Pan, J., Sharifzadeh, S., Kaissis, G., Tresp, V., et al. Relationformer: A unified framework for image-to-graph generation. In ECCV, pp.\ 422--439. Springer, 2022

  24. [32]

    S., Kersting, K., and Roth, S

    Sudhakaran, G., Dhami, D. S., Kersting, K., and Roth, S. Vision relation transformer for unbiased scene graph generation. In ICCV, pp.\ 21882--21893, 2023

  25. [33]

    Learning to compose dynamic tree structures for visual contexts

    Tang, K., Zhang, H., Wu, B., Luo, W., and Liu, W. Learning to compose dynamic tree structures for visual contexts. In CVPR, pp.\ 6619--6628, 2019

  26. [34]

    Unbiased scene graph generation from biased training

    Tang, K., Niu, Y., Huang, J., Shi, J., and Zhang, H. Unbiased scene graph generation from biased training. In CVPR, pp.\ 3713--3722, 2020

  27. [35]

    Llama 2: Open foundation and fine-tuned chat models

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

  28. [36]

    Attention is all you need

    Vaswani, A. Attention is all you need. NeurIPS, 2017

  29. [37]

    Object-aware distillation pyramid for open-vocabulary object detection

    Wang, L., Liu, Y., Du, P., Ding, Z., Liao, Y., Qi, Q., Chen, B., and Liu, S. Object-aware distillation pyramid for open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11186--11196, 2023

  30. [38]

    The all-seeing project v2: Towards general relation comprehension of the open world

    Wang, W., Ren, Y., Luo, H., Li, T., Yan, C., Chen, Z., Wang, W., Li, Q., Lu, L., Zhu, X., et al. The all-seeing project v2: Towards general relation comprehension of the open world. In ECCV, pp.\ 471--490. Springer, 2025

  31. [39]

    Wu, S., Zhang, W., Jin, S., Liu, W., and Loy, C. C. Aligning bag of regions for open-vocabulary object detection. In CVPR, pp.\ 15254--15264, 2023

  32. [40]

    B., and Fei - Fei, L

    Xu, D., Zhu, Y., Choy, C. B., and Fei - Fei, L. Scene graph generation by iterative message passing. In CVPR, pp.\ 3097--3106, 2017

  33. [41]

    and Kovashka, A

    Ye, K. and Kovashka, A. Linguistic structures as weak supervision for visual scene graph generation. In CVPR, pp.\ 8289--8299, 2021

  34. [42]

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

    Yu, Q., Li, J., Wu, Y., Tang, S., Ji, W., and Zhuang, Y. Visually-prompted language model for fine-grained scene graph generation in an open world. In ICCV, pp.\ 21560--21571, 2023

  35. [43]

    Zang, Y., Li, W., Zhou, K., Huang, C., and Loy, C. C. Open-vocabulary detr with conditional matching. In ECCV, pp.\ 106--122. Springer, 2022

  36. [44]

    Neural motifs: Scene graph parsing with global context

    Zellers, R., Yatskar, M., Thomson, S., and Choi, Y. Neural motifs: Scene graph parsing with global context. In CVPR, pp.\ 5831--5840, 2018

  37. [45]

    Zhang, Y., Pan, Y., Yao, T., Huang, R., Mei, T., and Chen, C. W. Learning to generate language-supervised and open-vocabulary scene graph using pre-trained visual-semantic space. In CVPR, pp.\ 2915--2924, 2023

  38. [46]

    Learning to generate scene graph from natural language supervision

    Zhong, Y., Shi, J., Yang, J., Xu, C., and Li, Y. Learning to generate scene graph from natural language supervision. In ICCV, pp.\ 1823--1834, 2021

  39. [47]

    Openpsg: Open-set panoptic scene graph generation via large multimodal models

    Zhou, Z., Zhu, Z., Caesar, H., and Shi, M. Openpsg: Open-set panoptic scene graph generation via large multimodal models. In ECCV, pp.\ 199--215. Springer, 2025

  40. [48]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.