Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

Semantic to Structure: Learning Structural Representations for Infringement Detection

T0 review · 3 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a contrastively trained structural representation, built from depth-conditioned diffusion pairs with LLM-rewritten captions, detects structural infringement that semantic embeddings miss, roughly tripling average…

desk verdict A useful task definition and a plausible synthesis pipeline, but the main benchmark is circular and the real-world set is too small to back the headline claim. read the letter →

arxiv 2502.07323 v1 pith:QGTGCV2J submitted 2025-02-11 cs.CV

Chuanwei Huang , Zexi Jia , Hongyan Fei , Yeshuang Zhu , Zhiqiang Yuan , Jinchao Zhang , Jie Zhou This is my paper · ORCID
classification cs.CV
keywords structuralinfringementrepresentationlearningcontrastivediffusionmodelsdepth-conditionedgenerationimageretrievalcopyrightprotectionAIGCsafety
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

Structural infringement, as defined here, is one image copying another image's geometry and element arrangement while changing the semantic content, a pattern that conventional semantic copy detection misses. The paper's claim is that a representation trained with contrastive learning on depth-conditioned diffusion pairs can isolate structure from semantics, letting a retrieval system flag structurally infringing images even when objects and text differ. To create training data, the authors extract a depth map with DPT, use an LLM to rewrite the caption so the subjects change, and ask ControlNet with SDXL to regenerate the image from the new caption under the depth condition. On two manually annotated benchmarks — SIA (513 synthetic pairs) and SIR (30 real pairs) — the learned descriptor beats DINOv2, MoCoV3, and SSCD, with SIA μAP rising from 0.129 to 0.365. If this holds, it gives creators and platforms a concrete way to catch a form of copying that current embeddings cannot see by construction.

What carries the argument

The load-bearing construction is the depth-conditioned diffusion pair. For a source image, DPT predicts a monocular depth map that acts as the structural condition; an LLM rewrites the caption to swap entities and attributes while keeping relationships; and ControlNet, on top of SDXL, generates a new image that follows the depth layout with the rewritten semantics. This pair is then fed through a MoCo-style contrastive loop with a ViT-L encoder fine-tuned via LoRA, using InfoNCE loss at temperature 0.2; the positive pair (source, regenerate) and negative pairs from other images force the embedding to preserve fine-grained geometric and positional information while discarding object identity and style.

What would settle it

Compute depth maps for the real SIR pairs and check whether human-rated infringing pairs actually share similar depth structure; if many are judged infringing despite divergent depth maps, depth-conditioned synthesis is not capturing the phenomenon. A direct experiment would take an image pair with identical composition but different camera roll or perspective, so depth maps diverge, and test whether the descriptor still flags them as structurally infringing — if it does, the representation uses cues beyond depth; if it does not, depth layout is the effective learned signal.

Watch

Extended reading notes

Core claim

The central discovery is that an 'image structural representation' can be learned without labels by treating a depth-conditioned diffusion regeneration as a positive view of the original image. The pipeline generates pairs with high structural similarity and low semantic similarity: DPT provides the depth map, an LLM rewrites the source caption to replace object categories and attributes, and ControlNet steers SDXL to produce a new image under that depth condition. Fine-tuning a ViT-L encoder with LoRA and InfoNCE loss pulls each original and its regenerate close in embedding space while pushing different images apart; the result is a descriptor that the paper shows retrieves structure matches rather than same-category matches. On the SIA test set, this descriptor reaches 0.365 μAP versus 0.129 for DINOv2 and 0.120 for MoCoV3, and on SIR it reaches 0.527 μAP versus 0.496 for MoCoV3, with qualitative retrievals showing layout-consistent images despite different content.

Load-bearing premise

The load-bearing premise is that the 'structural similarity' captured by sharing a DPT depth map, followed by diffusion regeneration, is the same phenomenon human judges call structural infringement; if depth maps encode only coarse scene layout while infringement lies in finer arrangement, pose, or composition, the training signal and the SIA test set could be measuring the wrong thing.

Editorial extensions

If this is right

  • If the claim holds, retrieval systems can surface structural infringement even after the copied image replaces objects, attributes, and style.
  • The depth-plus-LLM synthesis recipe gives a label-free way to generate structure-focused training pairs for other geometric or layout tasks.
  • The SIA and SIR benchmarks give future detectors a quantitative yardstick that semantic copy-detection metrics cannot provide.
  • Because the descriptor suppresses semantics, it can be combined with a semantic descriptor to cover both content copying and structure copying in one retrieval pipeline.

Reading between the lines

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

  • The paper leaves untested whether the learned cue is depth layout specifically or a broader arrangement signal; ablating the ControlNet condition (depth versus Canny edges versus pose) on SIR would show which structural cue the encoder actually latches onto.
  • With only 30 real query pairs, the SIR advantage over MoCoV3 is a thin slice; extending SIR to more styles and near-miss negatives would tell whether the depth-trained prior generalizes or fits layout-like cases too narrowly.
  • The same contrastive recipe could plausibly transfer to graphic design, UI screenshots, or 3D scenes, where composition and arrangement matter more than content, though the paper does not demonstrate this.
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

3 major / 7 minor

Summary. The paper introduces 'structural infringement' as images sharing high structural similarity but low semantic similarity, and proposes learning an Image Structural Representation by fine-tuning a DINOv2 backbone with LoRA under an InfoNCE contrastive loss. Training pairs are generated by a synthesis pipeline that extracts a DPT depth map from a source image, rewrites its caption with an LLM to change semantics, and generates a new image with SDXL plus ControlNet conditioned on the depth map. The authors construct two manually annotated benchmarks, SIA (513 synthetic pairs from WikiArt) and SIR (30 real-world pairs), and evaluate retrieval with μAP and mAP@k, reporting that their method outperforms DINOv2, MoCoV3, and SSCD on both.

Significance. The problem is timely and the proposed direction is plausible: if the representation truly captures structural similarity while ignoring semantics, it would be practically useful for AIGC copyright enforcement. I credit the authors with defining a concrete task, designing a simple and potentially reusable synthesis pipeline, manually annotating two test sets, and providing instructive qualitative evidence in Fig. 3 that DINO retrieves by object category while their method appears to respond to structure. The InfoNCE formulation is standard, and the method is in principle reproducible from the description, though some implementation details are missing. The main significance risk is evidential rather than mathematical: the SIA benchmark is generated by the same pipeline used for training, and the independent SIR benchmark is too small and lacks statistical support to fully confirm transfer to real structural infringement.

major comments (3)
  1. [Section III-A, Fig. 2] The SIA test set is circular with respect to the training data. SIA positives are produced by applying exactly the data synthesis pipeline in Fig. 2 (DPT depth map, SDXL+ControlNet, LLM caption rewrite) to WikiArt images, while the training set is produced by the same pipeline on COCO images. By construction, every SIA positive pair shares the DPT depth map used for generation, so the high SIA gains in Table I (Ours 0.365 μAP vs. DINOv2 0.129) may reflect matching depth-conditioned generation artifacts or coarse depth layout rather than structural infringement as a general visual phenomenon. The manual rating of pairs does not break this circularity because raters only see pairs that the pipeline already produced. To support the central claim, the authors should add a non-circular evaluation, for example: (i) a depth-map-only baseline using the same DINOv2 features computed on DPT depth images; (ii) SIA built from a different generation pipeline or a different control condition; or (iii) an analysis of how SIA positive pairs correlate with depth-map overlap.
  2. [Section III-E, Table II] The independent SIR result is too weakly evidenced, and the retrieval setup is under-specified. With only 30 real pairs, the reported differences in Table II (Ours 0.527 μAP, MoCoV3 0.496, DINOv2 0.461) are within a range that a few outlier pairs could explain; no error bars, confidence intervals, or significance tests are reported. The statement that 'performance rankings ... are roughly consistent' is therefore not statistically supported. The evaluation section also does not specify the query/gallery split or number of distractors for SIA, and for SIR it only says 20,000 images were added without describing their source, filtering, or whether they are known negatives. Please report bootstrap confidence intervals or permutation tests on SIR, describe both test-set retrieval protocols, and either collect more real pairs or qualify the SIR conclusions accordingly.
  3. [Section II, Section III-A] The central assumption that depth-conditioned generation preserves exactly the structure relevant to real structural infringement is not validated. DPT depth maps preserve coarse scene layout, but human judgments of structural infringement may depend on finer arrangement of elements, pose, and composition; if so, the training signal and the SIA test measure the wrong phenomenon. Concretely, the authors should measure the agreement between DPT-depth similarity and human infringement scores on SIR, and compare alternative ControlNet conditions (e.g., Canny edges, soft edges, or pose) to show that the depth condition is not merely the easiest artifact to match. This is a correctness-risk concern, not an internal inconsistency.
minor comments (7)
  1. [Section III-A] The placeholder 'dataset link' should be replaced with actual URLs for the SIA and SIR datasets, and any code or trained model should be released for reproducibility.
  2. [Section II] There are typographical issues, including 'Base on this' (should be 'Based on this') and 'a image' (should be 'an image'); the paper would also benefit from a consistent use of 'structural infringement' instead of occasionally abbreviating to 'image infringement' in the Index Terms.
  3. [Section III-C] The relationship between μAP and mAP@1/5/10 is not defined; please state how the mAP@k values are computed, since Tables I and II report both sets of metrics.
  4. [Fig. 3] The figure layout is ambiguous: the caption says 'for each pair' but the composition implies a fixed query column; please clarify the arrangement of query and retrieval images and what the numerical similarity scores correspond to.
  5. [Section III-A] Please report inter-annotator agreement for the manual filtering used to create SIA and SIR, e.g., Cohen's kappa or Krippendorff's alpha, since the ground truth is defined by an average threshold greater than 4.
  6. [Section II, Section III-B] The InfoNCE loss is sensitive to the number of negatives; please specify the batch size, whether a queue is used as in MoCo, the total number of training steps, and the exact DPT and ControlNet settings (model versions, resolution, number of inference steps).
  7. [Section III-E] The comparison would be strengthened by including a layout-representation baseline such as the method in [11], which is discussed in the introduction but not evaluated.

Circularity Check

1 steps flagged · score 4.0 of 10

SIA test set reuses the training synthesis pipeline, making the headline SIA gains partially circular; the small real-world SIR set provides only weak independent grounding.

  1. self definitional [Section III-A (SIA Dataset) and Section II (Methods), Fig. 2; Trainset in Section III-A]
    "SIA Dataset is constructed using a synthetic approach. Initially, 2,000 art images of various styles are randomly selected from WikiArt. Subsequently, we employ the data synthesis pipeline in Fig. 2 to generate infringing images. ... Pairs with average scores greater than 4 are retained, resulting in a testset of 513 pairs."

    The SIA test positives are generated by exactly the same DPT-depth + ControlNet/SDXL + LLM-rewrite pipeline (Fig. 2) used to create the training pairs from COCO. The training contrastive loss maximizes agreement between xsrc and xsyn from this pipeline, so the representation is fitted to recognize images produced by this generator. Evaluating on SIA therefore measures retrieval of images produced by the same fitted generator: the 'structural infringement' positives are, by construction, pairs that share the DPT depth map used for generation. The manual score >4 filter only confirms that the pipeline preserved depth structure, which is the same signal the InfoNCE objective is trained to encode.

full rationale

No self-citation chain or imported uniqueness theorem is load-bearing; the method is a straightforward contrastive fine-tuning of DINOv2 with LoRA. The circularity is concentrated in the SIA benchmark: both the training data and the SIA positive pairs come from the same SDXL+ControlNet depth-conditioned generation process. This makes the large SIA improvement (Ours 0.365 vs DINOv2 0.129) partially by construction, because the model is trained to match outputs of that exact generator. The SIR set is independently collected from real-world alleged infringement cases and provides genuine external grounding; however, with only 30 pairs and no error bars or significance testing, it cannot by itself resolve the circularity or fully support the strong claim of real-world structural infringement detection. The paper's statement that SIA/SIR ranking consistency 'indicating that our data synthesis pipeline can partially reflect real-world structural infringement phenomena' is an inference from a very small sample, not an independent validation. Overall, the central derivation is not mathematically circular, but the main evaluation is partially self-referential, so a score of 4 is appropriate.

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

The central claim depends on several hand-set choices (temperature tau equals 0.2, LoRA rank r equals 3, annotation threshold greater than 4, SIR gallery expansion by 20,000 images) and on untested domain assumptions about depth maps capturing infringement-relevant structure and LLM rewriting removing semantic similarity. No new physical entities are introduced; the only new conceptual object is the task definition and the learned representation.

free parameters (4)
  • InfoNCE temperature tau = 0.2
    Sets the contrastive sharpness during training; chosen by hand and not ablated.
  • LoRA rank r = 3
    Controls the capacity of the fine-tuning adapter; chosen by hand and not ablated.
  • SIA and SIR annotation threshold = average human rating greater than 4
    Pairs below this threshold are discarded from the test sets; this changes the difficulty and composition of the benchmarks.
  • SIR gallery expansion size = 20,000 images
    Added to enlarge the retrieval gallery for the small SIR set; affects reported precision and recall values.
assumptions (5)
  • domain assumption Depth maps preserve the structural and positional information relevant to infringement.
    Introduced in Section II: the DPT depth map is used as the ControlNet condition because it preserves rough geometric and positional information; this is assumed, not validated independently.
  • ad hoc to paper LLM caption rewriting reduces semantic similarity while leaving structure intact.
    Core to the synthesis pipeline; the paper provides qualitative examples but no quantitative measure of semantic similarity reduction.
  • domain assumption Contrastive InfoNCE training on source-synthetic pairs makes the encoder discard semantic content and retain structure.
    Inherited from MoCo-style contrastive learning, applied here to offline diffusion augmentations; the paper does not ablate whether semantic leakage remains.
  • domain assumption Retrieval cosine similarity is a valid operationalization of structural infringement.
    Evaluation follows the image copy detection paradigm, but no threshold calibration or human agreement analysis is given.
  • domain assumption Human annotations with average scores above 4 identify genuine infringement.
    Test set curation in Section III-A relies on subjective ratings; annotation details, number of annotators, and inter-annotator agreement are not reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic to Structure: Learning Structural Representations for Infringement Detection." pith.science (2026). https://pith.science/paper/QGTGCV2J

@misc{pith2026250207323,
  author       = {Pith},
  title        = {Pith review of: Semantic to Structure: Learning Structural Representations for Infringement Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGTGCV2J}},
  note         = {Machine review of arXiv:2502.07323}
}
read the original abstract

Structural information in images is crucial for aesthetic assessment, and it is widely recognized in the artistic field that imitating the structure of other works significantly infringes on creators' rights. The advancement of diffusion models has led to AI-generated content imitating artists' structural creations, yet effective detection methods are still lacking. In this paper, we define this phenomenon as "structural infringement" and propose a corresponding detection method. Additionally, we develop quantitative metrics and create manually annotated datasets for evaluation: the SIA dataset of synthesized data, and the SIR dataset of real data. Due to the current lack of datasets for structural infringement detection, we propose a new data synthesis strategy based on diffusion models and LLM, successfully training a structural infringement detection model. Experimental results show that our method can successfully detect structural infringements and achieve notable improvements on annotated test sets.

Figures

Figures reproduced from arXiv: 2502.07323 by the authors.

Figure 1
Figure 1. Structural infringement image pairs in SIR and SIA datasets. (a): The SIR dataset encompasses image pairs that exhibit structural infringement in the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Data synthesis pipeline. (a) Given a source image with a caption description, the depth map is first extracted using DPT to capture the main structural [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Top-1 retrieval image on SIA datasets using DINO and our proposed image structural representation. For each pair, the left image is the query, and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. From Imitation to Innovation: The Emergence of AI Unique Artistic Styles and the Challenge of Copyright Protection

    cs.CV 2025-07 conditional novelty 6.0 of 10

    ArtBulb uses style-description-guided multimodal clustering combined with MLLMs to judge whether AI-generated artworks have a unique, consistent, prompt-accurate style eligible for copyright protection.

  2. A Visual Leap in CLIP Compositionality Reasoning through Generation of Counterfactual Sets

    cs.CV 2025-07 conditional novelty 6.0 of 10

    Block-based diffusion generation of counterfactual image-text sets, combined with a set-aware loss, improves CLIP's compositional reasoning over several benchmarks, but the paper overstates one benchmark result and sh...

Reference graph

Works this paper leans on

26 extracted references · 13 canonical work pages · cited by 2 Pith papers

  1. [1]

    Image composition assess- ment with saliency-augmented multi-pattern pooling,

    B. Zhang, L. Niu, and L. Zhang, “Image composition assess- ment with saliency-augmented multi-pattern pooling,” arXiv preprint arXiv:2104.03133, 2021

    arXiv 2021

  2. [2]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840– 6851, 2020

    2020

  3. [3]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis,

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M ¨uller, J. Penna, and R. Rombach, “Sdxl: Improving latent diffusion models for high-resolution image synthesis,” arXiv preprint arXiv:2307.01952 , 2023

    arXiv 2023

  4. [4]

    Understanding and mitigating copying in diffusion models,

    G. Somepalli, V . Singla, M. Goldblum, J. Geiping, and T. Goldstein, “Understanding and mitigating copying in diffusion models,” Advances in Neural Information Processing Systems , vol. 36, pp. 47783–47803, 2023

    work page 2023

  5. [5]

    Diffusion art or digital forgery? investigating data replication in diffu- sion models,

    G. Somepalli, V . Singla, M. Goldblum, J. Geiping, and T. Goldstein, “Diffusion art or digital forgery? investigating data replication in diffu- sion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 6048–6058, 2023

    work page 2023

  6. [6]

    Fantastic copyrighted beasts and how (not) to generate them,

    L. He, Y . Huang, W. Shi, T. Xie, H. Liu, Y . Wang, L. Zettlemoyer, C. Zhang, D. Chen, and P. Henderson, “Fantastic copyrighted beasts and how (not) to generate them,” arXiv preprint arXiv:2406.14526 , 2024

    arXiv 2024

  7. [7]

    Evaluating and mitigating ip infringement in visual generative ai,

    Z. Wang, C. Chen, V . Sehwag, M. Pan, and L. Lyu, “Evaluating and mitigating ip infringement in visual generative ai,” arXiv preprint arXiv:2406.04662, 2024

    arXiv 2024

  8. [8]

    An empirical study of training self- supervised vision transformers,

    X. Chen, S. Xie, and K. He, “An empirical study of training self- supervised vision transformers,” in Proceedings of the IEEE/CVF in- ternational conference on computer vision , pp. 9640–9649, 2021

    2021

Show all 26 references
  1. [9]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. , “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  2. [10]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Proceedings of the IEEE/CVF international conference on computer vision, pp. 9650–9660, 2021

    2021

  3. [11]

    Self-supervised photographic image layout representation learning,

    Z. Zhao, P. Lu, X. Peng, and W. Guo, “Self-supervised photographic image layout representation learning,” arXiv preprint arXiv:2403.03740, 2024

    2024 arXiv

  4. [12]

    Hierarchical layout-aware graph convolutional network for unified aesthetics assessment,

    D. She, Y .-K. Lai, G. Yi, and K. Xu, “Hierarchical layout-aware graph convolutional network for unified aesthetics assessment,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8471–8480, 2021

    2021

  5. [13]

    Object-level attention for aesthetic rating distribution prediction,

    J. Hou, S. Yang, and W. Lin, “Object-level attention for aesthetic rating distribution prediction,” in Proceedings of the 28th ACM International Conference on Multimedia , pp. 816–824, 2020

    2020

  6. [14]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 3836–3847, 2023

    2023

  7. [15]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pp. 12179–12188, October 2021

    2021

  8. [16]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in International conference on machine learning , pp. 1597–1607, PMLR, 2020

    2020

  9. [17]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 9729–9738, 2020

    2020

  10. [18]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 , pp. 740–7...

    2014

  11. [19]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  12. [20]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685 , 2021

    2021 arXiv

  13. [21]

    Decoupled weight decay regularization,

    I. Loshchilov, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  14. [22]

    Billion-scale similarity search with gpus,

    J. Johnson, M. Douze, and H. J ´egou, “Billion-scale similarity search with gpus,” IEEE Transactions on Big Data , vol. 7, no. 3, pp. 535–547, 2019

    2019

  15. [23]

    A self- supervised descriptor for image copy detection,

    E. Pizzi, S. D. Roy, S. N. Ravindra, P. Goyal, and M. Douze, “A self- supervised descriptor for image copy detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pp. 14532–14542, 2022

    2022

  16. [24]

    A family of contextual measures of similarity between distributions with application to image retrieval,

    F. Perronnin, Y . Liu, and J.-M. Renders, “A family of contextual measures of similarity between distributions with application to image retrieval,” in 2009 IEEE Conference on computer vision and pattern recognition, pp. 2358–2365, IEEE, 2009

    2009

  17. [25]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 770–778, 2016

    2016

  18. [26]

    Aggregated residual transformations for deep neural networks,

    S. Xie, R. Girshick, P. Doll ´ar, Z. Tu, and K. He, “Aggregated residual transformations for deep neural networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 1492–1500, 2017

    2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.