Pith. sign in

REVIEW 3 major objections 4 minor 59 references

ProCrop: Learning Aesthetic Image Cropping from Professional Compositions

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ProCrop's central claim: retrieving and fusing features from compositionally similar professional photos outperforms existing cropping methods, and a 242K-image weakly labeled dataset lets it match fully supervised accuracy.

desk verdict A solid systems paper with a useful dataset, but the retrieval mechanism is under-evidenced and the weakly supervised loop has a self-training confound. read the letter →

arxiv 2505.22490 v1 pith:OGIYBEWR submitted 2025-05-28 cs.CV

Ke Zhang , Tianyu Ding , Jiachen Jiang , Tianyi Chen , Ilya Zharkov , Vishal M. Patel , Luming Liang This is my paper
classification cs.CV
keywords imagecroppingaestheticcompositionretrievalaugmentationweaklysupervisedlearningdatasetgenerationoutpaintingfeaturefusioncomposition-aware
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

This paper argues that aesthetic image cropping can be learned from professional photography without annotations on the reference images: given a query photo, ProCrop retrieves professional photos whose layout resembles the query's and fuses their features into the cropping network, so the model copies professional composition rules instead of relying on hand-crafted rules or labeled crop data. The second claim is that training data itself can be manufactured: outpainting 80,000 professional images into 242,000 synthetic 'uncropped' versions, with diverse crop proposals refined by the model itself, produces the largest public dataset for image cropping and lets a weakly supervised model match fully supervised performance. If both claims hold, cropping systems no longer depend on scarce human-rated crop annotations, and the same retrieve-then-fuse recipe could supply compositional knowledge to other layout-sensitive tasks. The reported results support the claims: on GAICv1 the method reaches 86.0 ACC5 and 94.5 ACC10, and on zero-shot SACD transfer it exceeds prior weakly supervised methods with IoU 0.7678.

What carries the argument

The load-bearing mechanism is retrieval-augmented feature fusion. The SAM encoder acts as a compositional feature extractor: its embeddings of the query image and of each entry in a professional database (the top 55,000 AVA images by aesthetic score) are matched to retrieve the ten most similar layouts, and those retrieved features are projected, cross-attended against the query's cDETR encoder features, and concatenated into the decoder input alongside BLIP multi-modal text embeddings. The decoder, built on parallel regression and classification heads with learnable anchors, outputs N crop boxes with aesthetic scores in one pass. The same machinery generates the training set: GPT-4 writes within- and beyond-image layout text, multi-scale SAM masks condition ControlNet outpainting, and iterative self-training with dynamic ranking of crop scores produces the pseudo-labels for the 242K-image CAD dataset.

What would settle it

Train the identical ProCrop model but replace similarity-based retrieval with random selection from the same professional database; if the accuracy gain over the no-retrieval baseline is fully preserved, the compositional-similarity mechanism is not what carries the improvement.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that compositional similarity, computed with the SAM encoder over a large unlabeled professional database, is a usable source of supervision for cropping: retrieving the top-K nearest professional images and feeding their features into the decoder through concatenation and cross-attention materially improves crop quality over the same model without retrieval. The same retrieval concept extends to data generation: GPT-4 dual-space descriptions and SAM masks drive ControlNet outpainting of professional images, and an iterative model-in-the-loop refinement converts the outpainted canvases into 242,000 diverse, weakly labeled image-crop pairs. With this composition-aware dataset, the paper reports that ProCrop surpasses prior weakly supervised methods on the unseen SACD test set and matches fully supervised accuracy, establishing the dataset as a substitute for expensive human crop ratings.

Load-bearing premise

The method rests on the assumption that two photos whose visual features are close, as measured by the segmentation model's encoder, also share a similar composition, so the professional photo's layout can teach the query photo's crop.

Editorial extensions

If this is right

  • On GAICv1's test set, ProCrop with retrieval from AVA reaches 86.0 ACC5 and 94.5 ACC10, beating all prior supervised methods listed in the paper's comparison.
  • On the unseen SACD test set, the weakly supervised model reaches IoU 0.7678 and Disp 0.0506 with three crop proposals, surpassing earlier weakly supervised methods and several fully supervised ones.
  • The retrieval database can be swapped at inference time without retraining: using Unsplash-lite instead of AVA during testing gives comparable results, which implies the model learns transferable composition guidance.
  • Performance scales with retrieval source quality and count: AVA's top 55,000 professional images outperform GAIC or CPC retrievals, and gains saturate near ten retrieved images.
  • The released 242,000-image CAD dataset becomes the largest public resource for aesthetic image cropping, providing roughly 1.9 million crop annotations for future training and benchmarking.

Reading between the lines

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

  • A quantitative retrieval-quality study (human ratings of whether SAM-nearest professional images share the query's layout) would pin down the mechanism and provide a calibration target for the SAM-based similarity.
  • Because retrieval databases can be swapped at inference time without retraining, a natural extension is a personalized cropping system whose retrieval index is the user's own curated photo set, letting the same model adapt its aesthetic preferences by changing the database alone.
  • If the composition-retrieval mechanism transfers, the same retrieve-then-fuse recipe could benefit other layout-sensitive tasks such as layout generation, poster design, or video thumbnail selection, where a professional corpus can supply composition templates without labels.
  • The synthetic pseudo-labels inherit the aesthetic biases of GPT-4 and ControlNet, so the claimed match with fully supervised training on CAD should be re-tested on image domains outside the nature-heavy Unsplash-lite source to see whether bias matters.
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 / 4 minor

Summary. ProCrop presents a retrieval-augmented aesthetic image cropping model. It uses SAM-encoder features to retrieve top-K compositionally similar professional images from an unlabeled database, fuses their features with the query's cDETR features and BLIP text embeddings, and decodes N crop proposals with aesthetic scores. It also introduces CAD, a 242K-image weakly annotated dataset generated by outpainting professional images with ControlNet and iteratively refining crop proposals via a model-in-the-loop pseudo-labeling process. Experiments report state-of-the-art supervised results on GAICv1/GAICv2 and strong zero-shot transfer to SACD, with ablations on retrieval sources and model components.

Significance. The claimed contributions—retrieval-based cropping without an annotated retrieval database and the largest publicly released cropping dataset—are potentially significant. The paper commits to releasing code and data, includes systematic ablations, and uses zero-shot SACD transfer as independent grounding. If the retrieval mechanism is confirmed, the approach is a genuine advance over existing supervised and weakly supervised methods. However, the current evidence does not yet establish that composition-aware SAM retrieval, rather than corpus quality, drives the gains; the pseudo-labeling loop introduces a confound in the retrieval ablation; and the multi-proposal evaluation protocol is underspecified.

major comments (3)
  1. [§3.2, §4.3, Table 5] The central mechanistic claim—that retrieving compositionally similar professional images via SAM features is what improves cropping—is not quantitatively supported. The only retrieval comparison is qualitative (Fig. 7), and Table 5 shows that retrieving from a 1K GAICv1 set degrades ACC5 from 0.815 to 0.805, while the 55K AVA set improves ACC5 to 0.860; this pattern is equally consistent with a large, professional, diverse corpus prior. Add controls that vary only the retrieval mechanism while holding the corpus fixed—e.g., random sampling from the AVA 55K set, CLIP/DreamSim nearest neighbors, and SAM nearest neighbors—and report downstream ACC5/ACC10 and SACD IoU/Disp for each. Without such controls, the paper's mechanism claim is not established.
  2. [§3.3, Table 6] The weakly supervised ablation has a self-training confound. The CAD pseudo-labels are generated by an iterative refinement in which a ProCrop model proposes crops and top-k proposals become pseudo-labels (§3.3). The paper does not state whether the 'w/o retrieval' and 'with retrieval' rows in Table 6 are trained on pseudo-labels produced by the full retrieval-augmented model or by the corresponding variant. If the former, the retrieval ablation cannot separate label source from retrieval mechanism. Please clarify the pseudo-label generation protocol for each row, and ideally regenerate pseudo-labels without retrieval for the no-retrieval condition.
  3. [§4.2, Table 4] The evaluation protocol for N>1 proposals is underspecified and is load-bearing for the claim that ProCrop outperforms single-output baselines. For N=2 and N=3, what is reported as IoU/Disp? If the max-IoU/min-Disp over the N proposals is taken, that protocol must be stated and matched to baselines; if the top-scoring proposal is used, the numbers should be defined accordingly. Also, the comparative numbers in Table 4 are borrowed from [15] rather than re-run under a common protocol, and no error bars or significance tests are reported anywhere, so 'significantly outperforms' is not statistically supported.
minor comments (4)
  1. [Fig. 6, Table 4] 'Horizonal' in Fig. 6 and 'Unspash' in Table 4 should be corrected to 'Horizontal' and 'Unsplash'.
  2. [§4.2] The definition of ACCK/N ('how many of the top-K predicted crops fall within this top-N MOS set') is inconsistent with the later use of ACC5/ACC10 as 'average top-k accuracy'; clarify the notation and how ACC5 relates to ACC1/5.
  3. [Fig. 9] The IOUi/Dispi labels in Fig. 9 are not defined; state that i indexes the number of proposals or the i-th proposal.
  4. [§4.3] The abstract claims that ProCrop trained on CAD 'matches fully supervised approaches,' but Table 4 compares only on zero-shot SACD; clarify whether this claim refers to those SACD baselines or to direct GAICv1/v2 comparisons.

Circularity Check

1 steps flagged · score 5.0 of 10

Weakly-supervised retrieval ablation is confounded by self-generated pseudo-labels; supervised and SACD results remain independent.

  1. fitted input called prediction [Section 3.3 (Composition-aware dataset generation) and Section 4.4 / Table 6 (weakly-supervised ablation)]
    "Section 3.3: 'We train a ProCrop model using these image-crop pairs. The model then enters an iterative cycle where it automatically generates crop proposals for each query image. These proposals undergo a curation process that selects a diverse set adhering to established aesthetic principles. During this iterative refinement process, we dynamically rank the aesthetic scores of the crop set. The top-k crops are then utilized as pseudo labels' ; Section 4.4: 'Results show that incorporating image retrieval leads to notable improvements in average IoU (0.7489 vs. 0.7103) and Disp (0.0555 vs."

    CAD's pseudo-labels are generated by the full retrieval-augmented ProCrop model itself: the model ranks its own crop proposals and the top-k become training labels. The Table 6 'w/o rtr.' baseline is then trained on those same retrieval-generated pseudo-labels and compared against the full retrieval model. This is not an independent test of the retrieval module: the with-retrieval model is, by construction, closer to the label generator, while the no-retrieval model is asked to reproduce crops whose production depended on retrieved features it never sees.

full rationale

ProCrop's central performance claims rest on three blocks: (i) supervised training on GAICv1/v2 and CPC with human-rated labels (Table 3), (ii) zero-shot transfer to SACD with human annotations (Table 4), and (iii) the weakly-supervised CAD pipeline. Blocks (i) and (ii) are externally anchored: their labels are independent human ratings, so the reported gains over prior methods are not constructed from the model's own outputs. I found no load-bearing self-citation chain: the cited 'line combination' premise [20,21] and the prior outpainting work [15] are not by the present authors, and no uniqueness theorem is imported. The one genuine circularity is confined to the weakly-supervised retrieval ablation: CAD pseudo-labels are generated by the full retrieval-augmented ProCrop (Section 3.3), and Table 6 then compares the full model against a no-retrieval variant trained on those same retrieval-generated labels. The retrieval gain in that table is therefore partly a measure of how well each variant can imitate the label-generating model, not an independent estimate of the retrieval module's contribution. The supervised retrieval-source ablation (Table 5) is cleaner and provides independent support for retrieval, which keeps the overall circularity partial rather than total. The absence of a random-retrieval or alternative-embedding control is an evidence gap about the mechanism, not a circularity, so it does not affect the score.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

No new physical or representational entities are posited; the contributions are an architecture and a dataset. The free parameters are standard ML hyperparameters, but several are not fully reported, notably the pseudo-label top-k and the N>1 evaluation protocol.

free parameters (6)
  • Retrieval count K = 10
    Number of retrieved professional images; selected after ablation in Figure 9 shows performance stabilizing around 10.
  • AVA aesthetic-score threshold = top 55,000 of 255,000
    Professional retrieval set uses top 55,000 AVA images by aesthetic score; threshold chosen by hand and affects database diversity and quality.
  • Number of crop proposals N = 90 (1-3 for reporting)
    Decoder produces 90 proposals, following prior work [18]; WS tables report top 1-3 proposals, and the evaluation protocol for N>1 is not fully specified.
  • Weakly supervised stage split = 100 epochs stage 1, 400 epochs stage 2
    Training on CAD is divided into initialization and pseudo-label refinement stages; split is a heuristic.
  • IoU equivalence threshold epsilon = 0.85
    Used to determine whether predicted and GT crops match in ACC metrics; inherited from prior work but affects all reported ACC numbers.
  • Pseudo-label top-k in iterative refinement
    The curation step selects 'top-k crops', but k is not reported; this is an unstated hyperparameter of the dataset generation loop.
assumptions (6)
  • domain assumption Compositional features can be characterized by line combinations in images.
    Invoked in Section 3.2 to justify retrieval with SAM features; no quantitative validation beyond qualitative figures.
  • domain assumption The original professional image is a high-quality crop of the outpainted canvas.
    Core labeling rule for CAD (Section 3.3); assumes outpainting preserves enough context that the original frame is an expert crop.
  • domain assumption Retrieved professional images from AVA top-rated set provide transferable aesthetic guidance for other images.
    The method's central mechanism; the paper provides ablations supporting it but no human study.
  • ad hoc to paper The iterative pseudo-label curation produces labels of sufficient quality to train the model without drifting.
    Self-training loop in Section 3.3; the paper does not measure label quality or drift.
  • domain assumption SAM and BLIP features used for retrieval and fusion encode composition rather than just semantics.
    Assumed throughout Sections 3.2 and D; only qualitative evidence is provided.
  • domain assumption GPT-4 and ControlNet produce coherent outpainted content that preserves composition.
    Dataset generation relies on this; Figure 4 shows examples but no systematic validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProCrop: Learning Aesthetic Image Cropping from Professional Compositions." pith.science (2026). https://pith.science/paper/OGIYBEWR

@misc{pith2026250522490,
  author       = {Pith},
  title        = {Pith review of: ProCrop: Learning Aesthetic Image Cropping from Professional Compositions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OGIYBEWR}},
  note         = {Machine review of arXiv:2505.22490}
}
read the original abstract

Image cropping is crucial for enhancing the visual appeal and narrative impact of photographs, yet existing rule-based and data-driven approaches often lack diversity or require annotated training data. We introduce ProCrop, a retrieval-based method that leverages professional photography to guide cropping decisions. By fusing features from professional photographs with those of the query image, ProCrop learns from professional compositions, significantly boosting performance. Additionally, we present a large-scale dataset of 242K weakly-annotated images, generated by out-painting professional images and iteratively refining diverse crop proposals. This composition-aware dataset generation offers diverse high-quality crop proposals guided by aesthetic principles and becomes the largest publicly available dataset for image cropping. Extensive experiments show that ProCrop significantly outperforms existing methods in both supervised and weakly-supervised settings. Notably, when trained on the new dataset, our ProCrop surpasses previous weakly-supervised methods and even matches fully supervised approaches. Both the code and dataset will be made publicly available to advance research in image aesthetics and composition analysis.

Figures

Figures reproduced from arXiv: 2505.22490 by the authors.

Figure 1
Figure 1. Overview of ProCrop’s retrieval-based aesthetic crop [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of ProCrop. Given an input image, ProCrop retrieves compositionally similar professional images and generates a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Composition-aware dataset generation. Professional images undergo three stages to create diverse image-crop pairs. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Outpainting results with three variations: (1) BLIP [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Distribution of compositional layouts across CAD. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of cropping results. Our approach preserves primary subjects (red boxes) while removing redundant [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Impact of retrieval count . We show the relationship [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Relationship between retrieved images and crop proposals. Unsplash-lite dataset is taken as the retrieve set for illustration. ranked by aesthetic scores for the top 55,000 images) and tested on the GAICv2 test set. During inference, the re￾trieval datasets include CP…
Figure 11
Figure 11. Figure 11: Illustration of text descriptions and corresponding outpainted results. The text descriptions are generated using BLIP, GPT-4 (within-image only), and GPT-4 (dual-space understanding), respectively [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Illustration of more out-painted examples: the visu￾alization of out-painted images and their diverse crop proposals. prompts used to generate within-image descriptions and dual-space understanding of image descriptions [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Diversity of retrieval datasets. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Visualization of SAM-extracted features (K-means [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Saliency visualization in complex scenarios. We extract [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 52 canonical work pages

  1. [15]

    Learning subject-aware cropping by outpainting professional photos

    James Hong, Lu Yuan, Micha ¨el Gharbi, Matthew Fisher, and Kayvon Fatahalian. Learning subject-aware cropping by outpainting professional photos. InProceedings of the AAAI Conference on Artificial Intelligence, pages 2175– 2183, 2024. 2, 4, 6, 7

    work page 2024

  2. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    arXiv

  3. [2]

    Retrieval-based language models and applications

    Akari Asai, Sewon Min, Zexuan Zhong, and Danqi Chen. Retrieval-based language models and applications. InPro- ceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 6: Tutorial Abstracts), pages 41–46, 2023. 3

    work page 2023

  4. [3]

    Retrieval-augmented diffusion models.Advances in Neural Information Processing Sys- tems, 35:15309–15324, 2022

    Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas M¨uller, and Bj ¨orn Ommer. Retrieval-augmented diffusion models.Advances in Neural Information Processing Sys- tems, 35:15309–15324, 2022

    work page 2022

  5. [4]

    Improving language models by retriev- ing from trillions of tokens

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retriev- ing from trillions of tokens. InInternational conference on machine learning, pages 2206–2240. PMLR, 2022. 2, 3

    work page 2022

  6. [5]

    Quantitative analysis of automatic image cropping algorithms:a dataset and comparative study

    Yi-Ling Chen, Tzu-Wei Huang, Kai-Han Chang, Yu-Chen Tsai, Hwann-Tzong Chen, and Bing-Yu Chen. Quantitative analysis of automatic image cropping algorithms:a dataset and comparative study. InIEEE WACV 2017, 2017. 2

    work page 2017

  7. [6]

    Quantitative analysis of automatic image cropping algorithms: A dataset and comparative study

    Yi-Ling Chen, Tzu-Wei Huang, Kai-Han Chang, Yu-Chen Tsai, Hwann-Tzong Chen, and Bing-Yu Chen. Quantitative analysis of automatic image cropping algorithms: A dataset and comparative study. In2017 IEEE winter conference on applications of computer vision (WACV), pages 226–234. IEEE, 2017. 2

    work page 2017

  8. [7]

    Learning to compose with professional pho- tographs on the web

    Yi-Ling Chen, Jan Klopp, Min Sun, Shao-Yi Chien, and Kwan-Liu Ma. Learning to compose with professional pho- tographs on the web. InProceedings of the 25th ACM inter- national conference on Multimedia, pages 37–45, 2017. 2, 7

    work page 2017

Show all 59 references
  1. [8]

    Reproducible scal- ing laws for contrastive language-image learning

    Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuh- mann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scal- ing laws for contrastive language-image learning. InPro- ceedings of the IEEE/CVF Conference on Computer V...

    2023

  2. [9]

    Au- tomatic image cropping using visual composition, boundary simplicity and content preservation models

    Chen Fang, Zhe Lin, Radomir Mech, and Xiaohui Shen. Au- tomatic image cropping using visual composition, boundary simplicity and content preservation models. InProceedings of the 22nd ACM international conference on Multimedia, pages 1105–1108, 2014. 1, 2, 5

    2014

  3. [10]

    Dream- sim: Learning new dimensions of human visual similarity using synthetic data.Advances in Neural Information Pro- cessing Systems, 36, 2024

    Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dream- sim: Learning new dimensions of human visual similarity using synthetic data.Advances in Neural Information Pro- cessing Systems, 36, 2024. 4, 7

    2024

  4. [11]

    Automatic image cropping for vi- sual aesthetic enhancement using deep neural networks and cascaded regression.IEEE Transactions on Multimedia, 20 (8):2073–2085, 2018

    Guanjun Guo, Hanzi Wang, Chunhua Shen, Yan Yan, and Hong-Yuan Mark Liao. Automatic image cropping for vi- sual aesthetic enhancement using deep neural networks and cascaded regression.IEEE Transactions on Multimedia, 20 (8):2073–2085, 2018. 1, 2

    2018

  5. [12]

    Retrieval augmented language model pre- training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval augmented language model pre- training. InInternational conference on machine learning, pages 3929–3938. PMLR, 2020. 2, 3

    2020

  6. [13]

    Salient-centeredness and saliency size in computational aes- thetics.ACM Transactions on Applied Perception, 20(2):1– 23, 2023

    Weng Khuan Hoh, Fang-Lue Zhang, and Neil A Dodgson. Salient-centeredness and saliency size in computational aes- thetics.ACM Transactions on Applied Perception, 20(2):1– 23, 2023. 3, 15

    2023

  7. [14]

    Composing photos like a photographer

    Chaoyi Hong, Shuaiyuan Du, Ke Xian, Hao Lu, Zhiguo Cao, and Weicai Zhong. Composing photos like a photographer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7057–7066, 2021. 1, 2, 7

    2021

  8. [16]

    Retrieval-augmented layout transformer for content-aware layout generation

    Daichi Horita, Naoto Inoue, Kotaro Kikuchi, Kota Yam- aguchi, and Kiyoharu Aizawa. Retrieval-augmented layout transformer for content-aware layout generation. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 67–76, 2024. 3, 4, 12, 15

    2024

  9. [17]

    Elasticsearch.https://www.elastic

    Huggingface. Elasticsearch.https://www.elastic. co/cn/elasticsearch, 2024. Accessed: 2024-11-14. 4, 13

    2024

  10. [18]

    Re- thinking image cropping: Exploring diverse compositions from global views

    Gengyun Jia, Huaibo Huang, Chaoyou Fu, and Ran He. Re- thinking image cropping: Exploring diverse compositions from global views. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 2446–2455, 2022. 4, 6, 7, 8, 12, 13

    2022

  11. [19]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 2, 3

    2023

  12. [20]

    Semantic line combination detector.arXiv preprint arXiv:2404.18399,

    Jinwon Ko, Dongkwon Jin, and Chang-Su Kim. Semantic line combination detector.arXiv preprint arXiv:2404.18399,

    arXiv

  13. [21]

    Photographic composition classification and dominant geo- metric element detection for outdoor scenes.Journal of Vi- 9 sual Communication and Image Representation, 55:91–105,

    Jun-Tae Lee, Han-Ul Kim, Chul Lee, and Chang-Su Kim. Photographic composition classification and dominant geo- metric element detection for outdoor scenes.Journal of Vi- 9 sual Communication and Image Representation, 55:91–105,

  14. [22]

    A2- rl: Aesthetics aware reinforcement learning for image crop- ping

    Debang Li, Huikai Wu, Junge Zhang, and Kaiqi Huang. A2- rl: Aesthetics aware reinforcement learning for image crop- ping. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 8193–8201, 2018. 2, 7

    2018

  15. [23]

    Learning to learn cropping models for different aspect ratio require- ments

    Debang Li, Junge Zhang, and Kaiqi Huang. Learning to learn cropping models for different aspect ratio require- ments. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 12685–12694,

  16. [24]

    Composing good shots by exploiting mutual relations

    Debang Li, Junge Zhang, Kaiqi Huang, and Ming-Hsuan Yang. Composing good shots by exploiting mutual relations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4213–4222, 2020. 1, 2, 7

    2020

  17. [25]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InInterna- tional conference on machine learning, pages 12888–12900. PMLR, 2022. 4, 15

    2022

  18. [26]

    Context-aware candidates for image cropping

    Tianpei Lian, Zhiguo Cao, Ke Xian, Zhiyu Pan, and Weicai Zhong. Context-aware candidates for image cropping. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1479–1483. IEEE, 2021. 1, 2

    2021

  19. [27]

    Optimizing photo composition

    Ligang Liu, Renjie Chen, Lior Wolf, and Daniel Cohen-Or. Optimizing photo composition. InComputer graphics fo- rum, pages 469–478. Wiley Online Library, 2010. 1, 2

    2010

  20. [28]

    Beyond image borders: Learn- ing feature extrapolation for unbounded image composition

    Xiaoyu Liu, Ming Liu, Junyi Li, Shuai Liu, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Beyond image borders: Learn- ing feature extrapolation for unbounded image composition. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13023–13032, 2023. 1, 2, 6, 7

    2023

  21. [29]

    Listwise view ranking for image cropping.IEEE Access, 7: 91904–91911, 2019

    Weirui Lu, Xiaofen Xing, Bolun Cai, and Xiangmin Xu. Listwise view ranking for image cropping.IEEE Access, 7: 91904–91911, 2019. 7

    2019

  22. [30]

    Conditional detr for fast training convergence

    Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 3651–3660, 2021. 4, 6

    2021

  23. [31]

    Ava: A large-scale database for aesthetic visual analysis

    Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408–2415, 2012. 2, 5, 6, 7

    2012

  24. [32]

    Learning to photograph: A compositional perspective.IEEE Transactions on Multime- dia, 15(5):1138–1151, 2013

    Bingbing Ni, Mengdi Xu, Bin Cheng, Meng Wang, Shuicheng Yan, and Qi Tian. Learning to photograph: A compositional perspective.IEEE Transactions on Multime- dia, 15(5):1138–1151, 2013. 1, 2

    2013

  25. [33]

    Sensation-based photo cropping

    Masashi Nishiyama, Takahiro Okabe, Yoichi Sato, and Imari Sato. Sensation-based photo cropping. InProceedings of the 17th ACM international conference on Multimedia, pages 669–672, 2009. 2

    2009

  26. [34]

    The role of image composition in image aesthetics

    Pere Obrador, Ludwig Schmidt-Hackenberg, and Nuria Oliver. The role of image composition in image aesthetics. In2010 IEEE International Conference on Image Process- ing, pages 3185–3188. IEEE, 2010. 1

    2010

  27. [35]

    Transview: Inside, outside, and across the cropping view boundaries

    Zhiyu Pan, Zhiguo Cao, Kewei Wang, Hao Lu, and Weicai Zhong. Transview: Inside, outside, and across the cropping view boundaries. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 4218–4227,

  28. [36]

    Retrieve-then-adapt: Example-based automatic generation for proportion-related infographics.IEEE Transactions on Visualization and Com- puter Graphics, 27(2):443–452, 2020

    Chunyao Qian, Shizhao Sun, Weiwei Cui, Jian-Guang Lou, Haidong Zhang, and Dongmei Zhang. Retrieve-then-adapt: Example-based automatic generation for proportion-related infographics.IEEE Transactions on Visualization and Com- puter Graphics, 27(2):443–452, 2020. 3

    2020

  29. [37]

    Highly accurate dichotomous im- age segmentation

    Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous im- age segmentation. InEuropean Conference on Computer Vi- sion, pages 38–56. Springer, 2022. 3

    2022

  30. [38]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, p...

    2021

  31. [39]

    A comparative study of image retargeting

    Michael Rubinstein, Diego Gutierrez, Olga Sorkine, and Ariel Shamir. A comparative study of image retargeting. In ACM SIGGRAPH Asia 2010 papers, pages 1–10. 2010. 2

    2010

  32. [40]

    Automatic image retargeting

    Vidya Setlur, Saeko Takagi, Ramesh Raskar, Michael Gle- icher, and Bruce Gooch. Automatic image retargeting. In Proceedings of the 4th international conference on Mobile and ubiquitous multimedia, pages 59–68, 2005. 2

    2005

  33. [41]

    Spatial-semantic collaborative cropping for user generated content

    Yukun Su, Yiwen Cao, Jingliang Deng, Fengyun Rao, and Qingyao Wu. Spatial-semantic collaborative cropping for user generated content. InProceedings of the AAAI Con- ference on Artificial Intelligence, pages 4988–4997, 2024. 6, 7

    2024

  34. [42]

    Image cropping with composition and saliency aware aes- thetic score map

    Yi Tu, Li Niu, Weijie Zhao, Dawei Cheng, and Liqing Zhang. Image cropping with composition and saliency aware aes- thetic score map. InProceedings of the AAAI conference on artificial intelligence, pages 12104–12111, 2020. 2, 7

    2020

  35. [43]

    Unsplash-lite dataset.https://unsplash

    Unsplash. Unsplash-lite dataset.https://unsplash. com/data, 2023. Accessed: 2023-12-15. 2, 6

    2023

  36. [44]

    Large-scale optimization of hierarchical features for saliency prediction in natural images

    Eleonora Vig, Michael Dorr, and David Cox. Large-scale optimization of hierarchical features for saliency prediction in natural images. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2798–2805,

  37. [45]

    Pos- sam: Panoptic open-vocabulary segment anything.arXiv preprint arXiv:2403.09620, 2024

    Vibashan VS, Shubhankar Borse, Hyojin Park, Debasmit Das, Vishal Patel, Munawar Hayat, and Fatih Porikli. Pos- sam: Panoptic open-vocabulary segment anything.arXiv preprint arXiv:2403.09620, 2024. 15

    2024 arXiv

  38. [46]

    Image cropping with spatial-aware feature and rank consistency

    Chao Wang, Li Niu, Bo Zhang, and Liqing Zhang. Image cropping with spatial-aware feature and rank consistency. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 10052–10061, 2023. 1, 6, 7

    2023

  39. [47]

    Deep cropping via at- tention box prediction and aesthetics assessment

    Wenguan Wang and Jianbing Shen. Deep cropping via at- tention box prediction and aesthetics assessment. InPro- ceedings of the IEEE international conference on computer vision, pages 2186–2194, 2017. 2

    2017

  40. [48]

    Good 10 view hunting: Learning photo composition from dense view pairs

    Zijun Wei, Jianming Zhang, Xiaohui Shen, Zhe Lin, Radomir Mech, Minh Hoai, and Dimitris Samaras. Good 10 view hunting: Learning photo composition from dense view pairs. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5437–5446, 2018. 2, 5, 7

    2018

  41. [49]

    Learning the change for automatic image cropping

    Jianzhou Yan, Stephen Lin, Sing Bing Kang, and Xiaoou Tang. Learning the change for automatic image cropping. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 971–978, 2013. 2

    2013

  42. [50]

    Focusing on your subject: Deep subject-aware image composition recommendation net- works.Computational Visual Media, 9(1):87–107, 2023

    Guo-Ye Yang, Wen-Yang Zhou, Yun Cai, Song-Hai Zhang, and Fang-Lue Zhang. Focusing on your subject: Deep subject-aware image composition recommendation net- works.Computational Visual Media, 9(1):87–107, 2023. 2, 5, 6

    2023

  43. [51]

    Reliable and efficient image cropping: A grid anchor based approach

    Hui Zeng, Lida Li, Zisheng Cao, and Lei Zhang. Reliable and efficient image cropping: A grid anchor based approach. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 5949–5957, 2019. 1, 2, 5, 7

    2019

  44. [52]

    Grid anchor based image cropping: A new benchmark and an efficient model.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1304–1319, 2020

    Hui Zeng, Lida Li, Zisheng Cao, and Lei Zhang. Grid anchor based image cropping: A new benchmark and an efficient model.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1304–1319, 2020. 1, 2, 5, 7

    2020

  45. [53]

    Human- centric image cropping with partition-aware and content- preserving features

    Bo Zhang, Li Niu, Xing Zhao, and Liqing Zhang. Human- centric image cropping with partition-aware and content- preserving features. InEuropean Conference on Computer Vision, pages 181–197. Springer, 2022. 6, 7

    2022

  46. [54]

    Detecting and removing visual distractors for video aesthetic enhancement.IEEE Transactions on Multimedia, 20(8):1987–1999, 2018

    Fang-Lue Zhang, Xian Wu, Rui-Long Li, Jue Wang, Zhao- Heng Zheng, and Shi-Min Hu. Detecting and removing visual distractors for video aesthetic enhancement.IEEE Transactions on Multimedia, 20(8):1987–1999, 2018. 3, 15

    1987

  47. [55]

    Weakly supervised photo cropping.IEEE Transactions on Multimedia, 16(1):94–107, 2013

    Luming Zhang, Mingli Song, Yi Yang, Qi Zhao, Chen Zhao, and Nicu Sebe. Weakly supervised photo cropping.IEEE Transactions on Multimedia, 16(1):94–107, 2013. 1, 2

    2013

  48. [56]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023. 2, 4

    2023

  49. [57]

    Auto cropping for digital photographs

    Mingju Zhang, Lei Zhang, Yanfeng Sun, Lin Feng, and Weiying Ma. Auto cropping for digital photographs. In2005 IEEE international conference on multimedia and expo, pages 4–pp. IEEE, 2005. 2

    2005

  50. [58]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 7

    2018

  51. [59]

    Composition-aware graphic layout gan for visual-textual presentation designs.arXiv preprint arXiv:2205.00303, 2022

    Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout gan for visual-textual presentation designs.arXiv preprint arXiv:2205.00303, 2022. 5, 7 11 ProCrop: Learning Aesthetic Image Cropping from Professional Compositions Supplem...

    2022 arXiv

Pith tools

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