Pith. sign in

REVIEW 3 major objections 5 minor 99 references

This paper tries to establish that assembly relationships can be read directly from part geometry, without semantic labels, and that this reading transfers to a vehicle type never seen during training.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 05:05 UTC pith:PE3W5OS6

load-bearing objection A sensible small-data assembly graph pipeline whose headline 'no semantic data' claim is contradicted by its own fine-tuned detector. the 3 major comments →

arxiv 2607.22325 v1 pith:PE3W5OS6 submitted 2026-07-24 cs.CV cs.RO

Geometric 2D Scene Graph Generation

Christoph Jahn , Urs Waldmann , Bastian Goldluecke This is my paper
classification cs.CV cs.RO
keywords scene graph generationassembly instructionsassembly roboticsadjacency matrix predictiongraph convolutional networkssmall dataset learninggeometric feature extractiontoy vehicle assembly
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Scene graph generation usually demands many labeled images and annotated subject-predicate-object triplets. This paper asks whether assembly structure can be separated from semantics: first decide which parts touch, then later assign labels and order. The authors train a three-stage model on disassembled parts from three toy vehicles—a train, a plane, and a boat—and report that it approximates the correct graph for an unseen fourth vehicle, a car. If the claim holds, assembly robots could derive assembly instructions directly from a photo of loose components, without knowing the components' names.

Core claim

The central claim is that the adjacency matrix—a binary table of which pairs of components are connected—can be predicted from geometry alone. The paper's pipeline first crops each component with an object detector and encodes its shape; a transformer with cross-attention and edge-augmented attention scores predicts the matrix in subgraphs; a Siamese network with graph-convolution message passing then consumes that matrix to predict relation labels (join, insert, screw), permutation constraints, and assembly order. On the study dataset, binary adjacency accuracy on the unseen car is 91.3% with MCC 0.76, while downstream sequence prediction drops to 82.6%. The authors present the work as a co

What carries the argument

The load-bearing object is the predicted adjacency matrix, a binary table saying for each ordered pair of detected components whether they connect. The transformer produces it from geometric feature vectors, augmenting ordinary cross-attention with a learned per-pair bias (an edge-augmented attention score) so pairwise geometry directly shapes the prediction. The resulting matrix is then fed to a Siamese graph-convolution network, whose message passing propagates information along exactly the predicted connections; all relation labels, permutation constraints, and assembly order predictions are conditioned on this same adjacency structure.

Load-bearing premise

The claim that the method needs no semantic data rests on treating the first-stage object detector as a neutral geometric cropper, but the paper fine-tunes that detector to recognize the six component classes—casing, wheel, screw, front, bottom, top—that also appear in the unseen test car, so generalization to truly unknown part categories is not tested.

What would settle it

Train a variant of the pipeline in which the detector is not fine-tuned on the six component classes, or give the test vehicle a part type never seen in training (e.g., a gear), and compare binary adjacency accuracy to the paper's reported numbers. If the score collapses to chance, the geometry-only claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • An assembly robot could photograph a set of loose parts and obtain the connectivity structure before knowing what any part is.
  • Scene graph prediction becomes feasible with very small datasets, because the model learns binary pairwise connectivity rather than dense triplet annotations.
  • The explicit separation of structure from semantics means the structural step can potentially transfer to new part shapes, as the unseen-car experiment suggests.
  • Because all later predictions are conditioned on one adjacency matrix, relation type, one-to-one constraints, and assembly order stay mutually consistent.
  • The predicted graph can be read directly as an assembly instruction, linking perception to planning.

Where Pith is reading between the lines

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

  • A stricter test of the 'no semantic data' claim would use an object detector that was not fine-tuned on the six vehicle part classes, or add a genuinely novel part type such as a gear to the test car; the current experiment cannot rule out semantic leakage from the detector.
  • If the geometric-first hypothesis is right, the same adjacency-matrix-first design should transfer to other assembly domains with scarce labels, such as furniture or electronics; that transfer is not demonstrated in the paper.
  • One testable extension is to replace the fine-tuned detector with an off-the-shelf segmentation model and check whether binary adjacency accuracy on the car survives; the paper's own confusion matrix shows attachment and housing parts are the main failure mode.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a three-step pipeline for generating geometric scene graphs from images of disassembled components. Step 1 uses a Faster R-CNN detector to obtain component crops, which are then converted to shape images and encoded with a pretrained ViT-B/32 (CLIP). Step 2 uses a transformer with cross-attention and an edge-augmented attention mechanism to predict a binary adjacency matrix, with subgraph sampling and a confidence-threshold post-processing step. Step 3 uses a Siamese network with GCN-based message passing to predict relation labels, permutations, and assembly sequences from the predicted adjacency matrix. The method is evaluated on a custom toy-vehicle dataset: training on train, plane, and boat graphs and evaluating on an unseen car graph. The paper reports accuracy, mean recall, and MCC for binary adjacency, label, permutation, and sequence prediction, and includes qualitative analyses of the unseen vehicle.

Significance. If the central claims were supported, the work would be relevant to assembly robotics and scene graph generation under data scarcity: a geometric-only, semantic-free pipeline that generalizes to unseen vehicle types from a very small labeled set would be a useful conceptual contribution. The paper has some strengths: it clearly separates structural prediction from semantic/procedural tasks, it uses synthetic primitive augmentations in a sensible attempt to regularize a transformer on tiny data, and it reports balanced metrics (MCC, mean recall) rather than accuracy alone. The conclusion also honestly acknowledges the severe limitations of the dataset size. However, the central claim that the approach does not rely on semantic data is internally inconsistent with the experimental design, and the evaluation is far too small and lacks baselines, error bars, and quantitative ablations to support the strong generalization claims. The contribution is therefore not established by the evidence presented.

major comments (3)
  1. [§3.1, §4.2, Abstract] The claim that the method 'does not rely on semantic data' and that 'object classes are not identified' is contradicted by §4.2, which states that Faster R-CNN is 'fine-tuned to detect six component classes: casing, wheel, screw, front, bottom and top.' The evaluation car is assembled from exactly these same six semantic classes. The class head of the detector is trained with semantic labels and is used to produce the crops that feed the geometric representation. In addition, the visual encoder is the CLIP ViT-B/32 model, pretrained on semantic image-text data. Consequently, the experiment never tests generalization to truly unknown part categories, and the central semantic-free claim is unsupported. To substantiate the claim, the detector would need to be class-agnostic (e.g., trained only as a proposal network) and the evaluation would need to include novel part categories not seen dur
  2. [§4.1, Table 1, §4.4] The evaluation rests on four graphs total: three training vehicles and one unseen vehicle. The unseen-vehicle results in Table 1 are weak for the tasks that matter (binary adjacency MCC 75.9, sequence MeanR 77.3), and no confidence intervals, standard deviations, or repeated runs are reported. The paper itself notes that performance is 'sensitive to the data sampled graph constellations' (§4.4), but does not quantify this sensitivity. Ablations are shown only as curves in Fig. 3 with no numerical table; the statement that 'without them recognition of the unknown vehicle fails' is not backed by quantitative evidence. Hyperparameters such as the confidence threshold P≈98%, focal-loss parameters α=0.55 and γ=2.0, and loss weights (1.0, 10^-2, 2.5·10^-4) are fixed without sensitivity analysis. No baseline comparisons are provided. These omissions make it impossible to assess whether the repo
  3. [§4.3, §4.4] The evaluation does not measure the actual end-to-end scene graph generation from images. §4.3 states: 'Assuming accurate detection via Faster R-CNN, we evaluate only the predicted graph structure.' Detection accuracy is never reported, and the effect of detection errors on the downstream adjacency and relation predictions is not assessed. Since the pipeline's first step is a learned detector, the absence of end-to-end metrics means the central claim of constructing scene graphs from images is only partially evaluated.
minor comments (5)
  1. [§3.4] There are numerous typographical artifacts from LaTeX ligatures: 'difficult', 'sufficient', 'Cofficient', 'pipleline'. These should be fixed in a revision.
  2. [§3.1] The phrase 'a Faster R-CNN extracts an individual image for each component' is imprecise: a detector predicts bounding boxes; it does not itself extract images. Use 'extracts a crop' or similar.
  3. [§3.4, Eq. (4)] The notation in Eq. (4) is inconsistent: if A_b is the normalized symmetric adjacency matrix, then L_b = D_b - A_b is not the standard normalized Laplacian. Please clarify the exact definition and how D_b is computed.
  4. [§3.3, Eq. (2)] The dimensions of the terms in Eq. (2) are not fully specified (e.g., W_S, W_O, W_H, U, V, and the concatenation). This makes the model description hard to reproduce. Please add a shape table or explicit dimensional annotations.
  5. [General] A data/code availability statement is missing. Given the custom dataset and the small scale of the experiments, releasing the dataset and implementation would substantially improve reproducibility.

Circularity Check

0 steps flagged

No load-bearing circularity: the held-out evaluation is clean. The 'no semantic data' claim is internally inconsistent with the six-class fine-tuned detector, but that is an evidence mismatch, not a derivation that reduces to its inputs.

full rationale

I walked the paper's derivation chain. The paper makes no analytic derivation that could reduce to its own inputs; all claims are empirical. The adjacency matrix is predicted by a transformer from ViT features of detected component crops, and label/permutation/sequence predictions are produced by a GCN-based Siamese network from that adjacency matrix. All predictions are compared against held-out ground truth for the car, a vehicle not used in training. The train/evaluation split is clean, and no equation or fitted parameter is simultaneously the model's input and its output. Equations (1)-(5) define standard attention, focal loss, Laplacian regularization, and group loss mechanisms, while Eq. (6) is the standard MCC metric; none of them encodes the evaluation result. The reference list contains no works by the present authors, so there is no self-citation chain and no imported uniqueness theorem. Hand-chosen hyperparameters (P about 98 percent, alpha=0.55, gamma=2.0, loss weights) are tuning choices, not circular predictions. The one serious concern is the semantic-leakage caveat: Section 3.1 claims the approach 'remains robust to unknown parts by avoiding semantic dependencies,' while Section 4.2 states 'Faster R-CNN is fine-tuned to detect six component classes: casing, wheel, screw, front, bottom and top,' and the evaluation car is assembled from those same six classes. The 'unseen vehicle' therefore tests new configurations of known categories, not genuinely unknown part types, so the central robustness claim is unsupported by the experimental design. This is a correctness/evidence problem, not circularity: the held-out graph predictions are not computed from, nor forced by, those class labels, and the reported evaluation still uses ground-truth adjacency matrices. I therefore list no circular step and assign score 1 for the internal-consistency caveat rather than 0.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

The paper introduces no new theoretical entities. Its contributions are architectural and empirical, so the main ledger items are hand-set hyperparameters and domain assumptions, plus the semantic-free assumption that is internally contradicted.

free parameters (5)
  • Confidence threshold P = ≈0.98
    Postprocessing hyperparameter used to merge subgraph adjacency predictions into a full n×n matrix; set without sensitivity analysis (Section 3.2).
  • Focal loss α = 0.55
    Positive-class weighting in focal loss (Eq. 3), hand-set.
  • Focal loss γ = 2.0
    Focusing parameter in focal loss (Eq. 3), hand-set.
  • Loss weights = 1.0 / 1e-2 / 2.5e-4
    Weights for focal, group, and Laplacian losses (Section 4.2), hand-set.
  • Transformer size and feature dim = 2 layers, 2 heads, hidden 72; 36-dim features
    Model capacity choices in Section 4.2; no ablation for these.
axioms (3)
  • domain assumption Faster R-CNN detection is accurate (no evaluation of detection errors).
    Section 4.3 states 'Assuming accurate detection via Faster R-CNN, we evaluate only the predicted graph structure.' End-to-end performance may be lower.
  • ad hoc to paper The six semantic classes used to fine-tune Faster R-CNN do not constitute semantic priors for graph generation.
    This is the only way to reconcile the abstract's 'does not rely on semantic data' with §4.2's fine-tuning on class labels; the paper never justifies this distinction.
  • domain assumption Primitive shape augmentation transfers geometric structure to unseen vehicle types.
    Section 4.1 and 4.4 attribute generalization to primitive augmentations, but there is no independent evidence that the synthetic shapes preserve the assembly-relevant geometry.

pith-pipeline@v1.3.0-alltime-deepseek · 11394 in / 11948 out tokens · 92441 ms · 2026-08-01T05:05:18.386595+00:00 · methodology

0 comments
read the original abstract

In production processes for consumer products, assembly instructions are essential not only for planning but also for executing the production process. Likewise in robotics, it is crucial for an assembly robot to understand how components fit together and can be assembled. To facilitate these tasks, we contribute a method for constructing scene graphs to represent and characterize assembly relationships between components. Our approach does not rely on semantic data and is capable of handling a very small dataset. To realize this, the output of a Faster R-CNN model is used to create geometric representations, which are then processed by a transformer architecture to generate an adjacency matrix. This matrix serves as input to a Siamese network that uses message passing based on an attentional graph convolutional network (aGCN) architecture to characterize the connections between the components. We validate our method on a study dataset of toy model components which can be assembled into transportation vehicles.

Figures

Figures reproduced from arXiv: 2607.22325 by Bastian Goldluecke, Christoph Jahn, Urs Waldmann.

Figure 1
Figure 1. Figure 1: To evaluate our approach, we conduct a case study (Fig. 1a–d) where the model is trained on [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Our three-step pipeline: First, component bounding boxes and geometric features are extracted [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quantitative results: Visualization of the mean recall (MeanR) on Validation I, Validation II and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results - The unknown vehicle is shown at the top left, followed at the bottom by its binary [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

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

99 extracted references · 17 linked inside Pith

  1. [1]

    and Choy, Li Fei-Fei , title =

    Krähenbühl, Danfei and Xu, Yuke and Zhu, Christopher B. and Choy, Li Fei-Fei , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  2. [2]

    Proceedings of the European Conference on Computer Vision (ECCV) , year =

    Yang, Jianwei and Lu, Jiasen and Lee, Stefan and Batra, Dhruv and Parikh, Devi , title =. Proceedings of the European Conference on Computer Vision (ECCV) , year =

  3. [3]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Cong, Yuren and Yang, Michael Ying and Rosenhahn, Bodo , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  4. [4]

    and Kaiser, Łukasz and Polosukhin, Illia , title =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Łukasz and Polosukhin, Illia , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  5. [5]

    and Bernstein, Michael and Fei-Fei, Li , title =

    Johnson, Justin and Krishna, Ranjay and Stark, Michael and Li, Li-Jia and Shamma, David A. and Bernstein, Michael and Fei-Fei, Li , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  6. [6]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Sadeghi, Mohammad Amin and Farhadi, Ali , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  7. [7]

    Proceedings of the European Conference on Computer Vision (ECCV) , year =

    Lu, Cewu and Krishna, Ranjay and Bernstein, Michael and Fei-Fei, Li , title =. Proceedings of the European Conference on Computer Vision (ECCV) , year =

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Wald, Johanna and Dhamo, Helisa and Navab, Nassir and Tombari, Federico , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  9. [9]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Dhamo, Helisa and Manhardt, Fabian and Navab, Nassir and Tombari, Federico , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  10. [10]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , year =

    Li, Yikang and Ouyang, Wanli and Zhou, Bolei and Wang, Kun and Wang, Xiaogang , title =. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , year =

  11. [11]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Zellers, Rowan and Yatskar, Mark and Thomson, Sam and Choi, Yejin , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  12. [12]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

    Li, Rongjie and Zhang, Songyang and He, Xuming , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

  13. [14]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Lin, Xin and Ding, Changxing and Zeng, Jinquan and Tao, Dacheng , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  14. [16]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Deformable DETR: Deformable transformers for end-to-end object detection , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  15. [17]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Attentive Relational Networks for Mapping Images to Scene Graphs , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  16. [18]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Bipartite Graph Network with Adaptive Message Passing for Unbiased Scene Graph Generation , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2021 , organization =

    2021

  17. [20]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Sarkar, Sayan Deb and Miksik, Ondrej and Pollefeys, Marc and Barath, Daniel and Armeni, Iro , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  18. [21]

    arXiv , year =

    SGTR+: End-to-End Scene Graph Generation with Transformer , author =. arXiv , year =

  19. [23]

    2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Grid: Scene-graph-based instruction-driven robotic task planning , author=. 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2024 , organization=

    2024

  20. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Learning to compose dynamic tree structures for visual contexts , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  21. [26]

    Journal of Machine Learning Research , volume=

    Rademacher and Gaussian Complexities: Risk Bounds and Structural Results , author=. Journal of Machine Learning Research , volume=

  22. [27]

    1998 , publisher=

    Statistical Learning Theory , author=. 1998 , publisher=

    1998

  23. [28]

    2012 , publisher=

    Foundations of Machine Learning , author=. 2012 , publisher=

    2012

  24. [32]

    Advances in Neural Information Processing Systems , volume =

    Do vision transformers see like convolutional neural networks? , author =. Advances in Neural Information Processing Systems , volume =

  25. [33]

    International Conference on Machine Learning (ICML) , year =

    Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks , author =. International Conference on Machine Learning (ICML) , year =

  26. [35]

    arXiv preprint arXiv:1611.01144 , year =

    Categorical Reparameterization with Gumbel-Softmax , author =. arXiv preprint arXiv:1611.01144 , year =

    Pith/arXiv arXiv

  27. [36]

    NeurIPS 2020 , year =

    Calibrating Deep Neural Networks using Focal Loss , author =. NeurIPS 2020 , year =

    2020

  28. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Shangwen Guo and Zhipeng Luo and Weili Guan and Jingen Liu and Tao Yu and Limin Wang and Bingbing Ni , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  29. [38]

    Computer Methods and Programs in Biomedicine , volume=

    Evaluating classification models in the context of imbalanced datasets: the application of Matthews correlation coefficient , author=. Computer Methods and Programs in Biomedicine , volume=. 2021 , publisher=

    2021

  30. [39]

    arXiv preprint arXiv:2002.05709 , year=

    A Simple Framework for Contrastive Learning of Visual Representations , author=. arXiv preprint arXiv:2002.05709 , year=

    Pith/arXiv arXiv 2002

  31. [40]

    arXiv preprint arXiv:1703.03400 , year=

    Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks , author=. arXiv preprint arXiv:1703.03400 , year=

    Pith/arXiv arXiv

  32. [41]

    arXiv preprint arXiv:2207.14529 , year=

    The Effects of Data Quality on Machine Learning Performance , author=. arXiv preprint arXiv:2207.14529 , year=

    Pith/arXiv arXiv

  33. [42]

    Information and Software Technology , volume=

    A survey on dataset quality in machine learning , author=. Information and Software Technology , volume=. 2023 , publisher=

    2023

  34. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Gong, Chengyue and Wang, Dilin and Li, Meng and Chandra, Vikas and Liu, Qiang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2021 , pages =

    2021

  35. [44]

    International Conference on Learning Representations , year =

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author =. International Conference on Learning Representations , year =

  36. [45]

    Advances in Neural Information Processing Systems , volume =

    Adversarial Examples Are Not Bugs, They Are Features , author =. Advances in Neural Information Processing Systems , volume =

  37. [46]

    2020 , eprint =

    Time for a Background Check! Uncovering the Impact of Background Features on Deep Neural Networks , author =. 2020 , eprint =

    2020

  38. [47]

    European Conference on Computer Vision (ECCV) , year =

    Microsoft COCO: Common Objects in Context , author =. European Conference on Computer Vision (ECCV) , year =

  39. [50]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Graph Contrastive Learning with Augmentations , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  40. [53]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Supervised Contrastive Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  41. [54]

    CVPR Workshops , year =

    Deep Graph Laplacian Regularization for Robust Denoising of Real Images , author =. CVPR Workshops , year =

  42. [55]

    Journal of Machine Learning Research , volume =

    Manifold Regularization: A Geometric Framework for Learning from Labeled and Unlabeled Examples , author =. Journal of Machine Learning Research , volume =. 2006 , publisher =

    2006

  43. [56]

    Proceedings of the European Conference on Computer Vision (ECCV) , year =

    Ismail Elezi and Sebastiano Vascon and Alessandro Torcinovich and Marcello Pelillo and Laura Leal-Taixé , title =. Proceedings of the European Conference on Computer Vision (ECCV) , year =

  44. [57]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Chen, Xinlei and He, Kaiming , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2021 , pages =

    2021

  45. [58]

    Sensors , volume=

    Graph Representation Learning and Its Applications: A Survey , author=. Sensors , volume=. 2023 , publisher=

    2023

  46. [59]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages=

    DeepSetNet: Predicting Sets with Deep Neural Networks , author=. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages=

  47. [60]

    Proceedings of the 37th International Conference on Machine Learning (ICML) , volume=

    Learn to Predict Sets Using Feed-Forward Neural Networks , author=. Proceedings of the 37th International Conference on Machine Learning (ICML) , volume=. 2020 , publisher=

    2020

  48. [61]

    International Conference on Learning Representations (ICLR) , year=

    DropEdge: Towards Deep Graph Convolutional Networks on Node Classification , author=. International Conference on Learning Representations (ICLR) , year=

  49. [63]

    Belkin, M., Niyogi, P., and Sindhwani, V. (2006). Manifold regularization: A geometric framework for learning from labeled and unlabeled examples. In Journal of Machine Learning Research , volume 7, pages 2399--2434. JMLR.org

    2006

  50. [64]

    and He, K

    Chen, X. and He, K. (2021). Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 15750--15758

    2021

  51. [65]

    Chicco, D. (2021). Evaluating classification models in the context of imbalanced datasets: the application of matthews correlation coefficient. Computer Methods and Programs in Biomedicine , 211:106447

    2021

  52. [66]

    Y., and Rosenhahn, B

    Cong, Y., Yang, M. Y., and Rosenhahn, B. (2023). Reltr: Relation transformer for scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages 1234--1243

    2023

  53. [67]

    Correia, G., Ribeiro, B., et al. (2023). Sparse transformers for graph representation learning. arXiv preprint arXiv:2301.10980

    Pith/arXiv arXiv 2023

  54. [68]

    Dhamo, H., Manhardt, F., Navab, N., and Tombari, F. (2021). Graph-to-3d: End-to-end generation and manipulation of 3d scenes using scene graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages 10134--10143

    2021

  55. [69]

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. (2021). An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations

    2021

  56. [70]

    Elezi, I., Vascon, S., Torcinovich, A., Pelillo, M., and Leal-Taixé, L. (2020). The group loss for deep metric learning. In Proceedings of the European Conference on Computer Vision (ECCV) . Springer, Cham

    2020

  57. [71]

    Fang, Y., Li, Y., Wang, Y., Zhang, Y., and Liu, S. (2023). Dropmessage: Unifying message passing and regularization for graph neural networks. arXiv preprint arXiv:2302.08772

    Pith/arXiv arXiv 2023

  58. [72]

    Gong, C., Wang, D., Li, M., Chandra, V., and Liu, Q. (2021). Keepaugment: A simple information-preserving data augmentation approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 1055--1064

    2021

  59. [73]

    T., Jeon, H

    Hoang, V. T., Jeon, H. J., You, E. S., Yoon, Y., Jung, S., and Lee, O. J. (2023). Graph representation learning and its applications: A survey. Sensors , 23(8):4168

    2023

  60. [74]

    Huang, C., Liu, Z., Qiu, X., Song, L., Liu, Q., and Zhou, J. (2021). Permutation equivariance of transformers and its applications. arXiv preprint arXiv:2106.07479

    Pith/arXiv arXiv 2021

  61. [75]

    S., Zaki, M

    Hussain, M. S., Zaki, M. J., and Subramanian, D. (2021). Global self-attention as a replacement for graph convolution. arXiv preprint arXiv:2108.03348

    Pith/arXiv arXiv 2021

  62. [76]

    Ilyas, A., Santurkar, S., Tsipras, D., Engstrom, L., Tran, B., and Madry, A. (2019). Adversarial examples are not bugs, they are features. In Advances in Neural Information Processing Systems , volume 32

    2019

  63. [77]

    A., Bernstein, M., and Fei-Fei, L

    Johnson, J., Krishna, R., Stark, M., Li, L.-J., Shamma, D. A., Bernstein, M., and Fei-Fei, L. (2015). Image retrieval using scene graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 3668--3678

    2015

  64. [78]

    Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., and Krishnan, D. (2020). Supervised contrastive learning. In Advances in Neural Information Processing Systems (NeurIPS)

    2020

  65. [79]

    Kim, S., Choi, K., Lee, J., Kim, D., and Kim, S. (2024). Depth-wise convolutions in vision transformers for efficient training on small datasets. arXiv preprint arXiv:2403.19394

    Pith/arXiv arXiv 2024

  66. [80]

    B., and Choy, L

    Krähenbühl, D., Xu, Y., Zhu, C. B., and Choy, L. F.-F. (2017). Scene graph generation by iterative message passing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 3097--3106

    2017

  67. [81]

    Lee, J., Lee, H., Kim, M., and Lee, H. Y. (2019). Set transformer: A framework for attention-based permutation-invariant neural networks. In International Conference on Machine Learning (ICML)

    2019

  68. [82]

    Li, R., Zhang, S., and He, X. (2023). Sgtr: End-to-end scene graph generation with transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence , 45(2):2150--2163

    2023

  69. [83]

    Li, R., Zhang, S., Wan, B., and He, X. (2021). Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 10174--10184. IEEE

    2021

  70. [84]

    Li, Y., Ouyang, W., Zhou, B., Wang, K., and Wang, X. (2017). Scene graph generation from objects, phrases and region captions. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages 1261--1269

    2017

  71. [85]

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll \'a r, P., and Zitnick, C. L. (2014). Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV) , pages 740--755. Springer

    2014

  72. [86]

    Lin, X., Ding, C., Zeng, J., and Tao, D. (2020). Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 3746--3753

    2020

  73. [87]

    Liu, J., Xu, M., Yan, X., Chen, W., Liu, H., Zhang, R., Li, Y., Zhang, L., and Li, Y. (2024). Joint generative modeling of scene graphs and images via diffusion models. arXiv preprint arXiv:2401.01130

    Pith/arXiv arXiv 2024

  74. [88]

    Liu, Y., Yang, X., Zhou, S., and Liu, X. (2022). Simple contrastive graph clustering. arXiv preprint arXiv:2205.07865

    Pith/arXiv arXiv 2022

  75. [89]

    Lu, C., Krishna, R., Bernstein, M., and Fei-Fei, L. (2016). Visual relationship detection with language priors. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 852--869

    2016

  76. [90]

    Mao, C., Wang, Z., et al. (2022). Graph partition neural networks for semi-supervised classification. arXiv preprint arXiv:2206.07670

    Pith/arXiv arXiv 2022

  77. [91]

    Ni, Z., Deng, X., Tai, C., Zhu, X., Xie, Q., Huang, W., Wu, X., and Zeng, L. (2024). Grid: Scene-graph-based instruction-driven robotic task planning. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 13765--13772. IEEE

    2024

  78. [92]

    Raghu, M., Unterthiner, T., Kornblith, S., Zhang, C., and Dosovitskiy, A. (2021). Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems , 34:12116--12128

    2021

  79. [93]

    Rezatofighi, H., Boloor, A., Kumar, A., Milan, A., Dick, A., and Reid, I. (2020). Learn to predict sets using feed-forward neural networks. In Proceedings of the 37th International Conference on Machine Learning (ICML) , volume 119, pages 8114--8124. PMLR

    2020

  80. [94]

    Rezatofighi, H., Tsoi, N., Gwak, J., Sadeghian, A., Reid, I., and Savarese, S. (2017). Deepsetnet: Predicting sets with deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages 5227--5236

    2017

Showing first 80 references.