Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Retrieval Augmented Image Harmonization

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that retrieving a single illumination-matched reference image can make image harmonization substantially more accurate, and reports new best results in both the with-reference and no-reference settings.

desk verdict Raiha has a solid retrieval-based idea but the SOTA claim is undercut by an evaluation that gives the method target-image references in training and gives baselines no references at all. read the letter →

arxiv 2412.13916 v1 pith:H57YOZ4D submitted 2024-12-18 cs.CV

Haolin Wang , Ming Liu , Zifei Yan , Chao Zhou , Longan Xiao , Wangmeng Zuo This is my paper · ORCID
classification cs.CV
keywords imageharmonizationretrieval-augmentedmethodcross-attentionilluminationconsistencyRetinextheorydensevisualfeaturessemantic-guidedfusionbenchmarkRAHarmony4
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

Image harmonization adjusts a pasted object's brightness and color so it blends into a new background. This paper argues that the usual approach, which relies on similar content already present in the background, is frequently ill-posed because such content is often absent. The proposed Raiha framework instead retrieves one reference image containing an object similar to the foreground and photographed under illumination matching the background, then feeds that reference into a U-Net that treats it as an extension of the background. On the iHarmony4 benchmark Raiha reports an average MSE of 15.60 versus 16.55 for the prior best method HDNet when no reference is used, and on a new retrieval-augmented benchmark, RAHarmony4, it reports 15.86 versus 18.72. If these numbers hold, Raiha is the current state of the art in both settings and a step toward making harmonization work when the background gives no clue.

What carries the argument

The central machinery is the harmonization-oriented retrieval pipeline combined with a Semantic-Guided Fusion (SGF) module. The retrieval pipeline uses a pre-trained denoising vision transformer (DVT) to extract dense patch features for content matching, and HSV histograms of image patches for illumination matching, with Retinex theory used to justify that content-plus-appearance determines illumination. The SGF module takes the DVT features ($c_f$, $c_b$, $c_r$) of foreground, background, and reference, concatenates them with the encoder features along the channel dimension, and computes the cross-attention map as $A = \phi([e_f, c_f]_c) \times \phi([[e_b, e_r]_s, [c_b, c_r]_s]_c)$, so that attention is restricted to regions semantically related to the foreground. A data augmentation strategy converts the ground-truth target image into additional references by random cropping, flipping, and resizing, which lets the network learn to use references without overfitting to the exact target.

What would settle it

Recompute RAHarmony4 metrics with a gallery that omits the query's own ground-truth target image, using only the other 7,403 training images as candidates. If Raiha's average MSE advantage over HDNet at 15.86 versus 18.72 disappears or shrinks below the non-reference gap when the exact target is excluded, then the reported retrieval gain rests on retrieving the ground truth rather than on the illumination-consistency criteria.

Watch

Extended reading notes

Core claim

The paper's central claim is that image harmonization benefits from an external reference chosen by a retrieval step tailored to the task, and that the same network, trained on a mix of reference and non-reference inputs, also improves the ordinary non-reference setting. The retrieval step has two requirements: the reference must contain an object whose dense visual features match the foreground (cosine similarity above 0.7), and the reference must share illumination with the background, judged by matching DVT content features and HSV patch histograms (threshold 0.9), justified by Retinex theory that appearance separates into reflectance and illumination. A Semantic-Guided Fusion module then concatenates the DVT features with encoder features when forming the cross-attention map, restricting attention to semantically relevant regions of the reference and background. Training alternates between retrieved references and augmented versions of the target image, and the authors report that Raiha outperforms all compared methods on iHarmony4 under the non-reference setting and on the newly constructed RAHarmony4 under the retrieval-augmented setting.

Load-bearing premise

The pipeline assumes that matching DVT content features and HSV patch histograms above fixed thresholds reliably identifies reference images whose illumination matches the background, and that the RAHarmony4 gallery of ground-truth target images is a realistic stand-in for the references a deployment would actually retrieve; if either fails, the reported gains shrink or reverse.

Editorial extensions

If this is right

  • If Raiha's reported numbers hold, harmonic harmonization no longer requires similar content to be present in the background; a single external reference suffices to reduce ill-posedness.
  • The same network works in both retrieval-augmented and non-reference settings because training randomly samples both, so deployment can fall back to no reference when retrieval fails.
  • The Semantic-Guided Fusion module shows that content priors from dense visual features can suppress attention to irrelevant regions when using an external reference.
  • The RAHarmony4 benchmark, built from iHarmony4 training targets, provides a testbed for retrieval-augmented harmonization that future methods can directly compare against.
  • Illumination consistency is an essential criterion for retrieval: the paper's ablations show that as the illumination threshold $\epsilon_a$ is relaxed, performance degrades and can fall below the non-reference setting.

Reading between the lines

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

  • Beyond the paper: the hand-set thresholds (0.7 and 0.9) in the retrieval criteria could be replaced by a learned joint embedding of foreground content and background illumination, enabling end-to-end training of retrieval and harmonization.
  • Beyond the paper: the RAHarmony4 gallery consists of ground-truth target images, an idealized reference source; in a realistic deployment with unrelated gallery images, the retrieved references may be less matching, so the reported gains could shrink—a scenario the paper does not test.
  • Beyond the paper: the same retrieval-plus-semantic-attention recipe could transfer to other image-editing tasks that need external guidance, such as relighting, shadow removal, or object insertion, though this is speculative.
  • Beyond the paper: reusing the DVT features for both retrieval and fusion suggests that content features are a reusable asset across the search and generation stages; a single-stage system could be trained to jointly retrieve and harmonize.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. Raiha is a retrieval-augmented image harmonization framework. It first retrieves a reference image from a gallery by matching foreground content via DVT dense features and background illumination via HSV histograms, then feeds the composite and reference through a U-Net equipped with a Semantic-Guided Fusion module that uses the DVT features to constrain cross-attention. For training, the authors combine retrieved references with augmented references created by randomly cropping, flipping, and resizing the ground-truth target image, and they randomly mix non-reference and retrieval-augmented iterations. Evaluations are on iHarmony4 (non-reference) and a newly constructed RAHarmony4 benchmark (retrieval-augmented), where the paper claims state-of-the-art MSE/PSNR in both settings.

Significance. If the comparisons were valid, the contribution would be substantive: an external-reference mechanism for harmonization, a new benchmark, and a reusable retrieval criterion. The paper also has useful ideas, notably reusing DVT features for both retrieval and attention guidance and using target-derived references to teach reference utilization. These strengths are undermined, however, by the experimental protocol: the non-reference comparison is contaminated by training-time target-derived references, the retrieval-augmented comparison is asymmetric, and the benchmark gallery is idealized. The fixes are within the scope of a revision, but the current evidence does not establish the central state-of-the-art claim.

major comments (4)
  1. [§III-C(c), Table I] The non-reference comparison is not a controlled test of the architecture. In Sec. III-C(c), the training protocol creates 'augmented references' by randomly cropping, flipping, and resizing the ground-truth target image, and Sec. III-C(d) says non-reference and retrieval-augmented settings are randomly sampled each training iteration. Thus the model evaluated in Table I is trained with an auxiliary reference stream that is derived from the answer image, while none of the baselines receive this privileged signal. The reported margin (Raiha 15.60 vs. HDNet 16.55 MSE) could therefore reflect training-time exposure to target appearance rather than a better harmonization architecture. Please provide a non-reference ablation trained without any target-derived references and report its iHarmony4 numbers.
  2. [§IV-C, Table II] The retrieval-augmented comparison is also asymmetric: Raiha receives a retrieved reference at inference, while HDNet, PCT-Net, and the other baselines receive only the composite image. Table II therefore conflates an additional input modality with model quality and does not support the claim that Raiha is the best harmonization network. To make the comparison meaningful, either give the baselines the same retrieved reference (for example by concatenating it as in Eq. (7)) or evaluate a completely non-reference version of Raiha on the same inputs. The 'Raiha non-reference' row in Table III is not a clean non-reference model because it shares the same training protocol with target-derived references.
  3. [§IV-A, RAHarmony4 construction] The RAHarmony4 gallery is constructed from iHarmony4 target images, and the testing set is a subset of those images. Because every reference is a ground-truth target image, the retrieval-augmented evaluation is performed in an idealized setting where the gallery contains perfectly harmonized images rather than arbitrary unprocessed images. Please clarify whether the gallery overlaps with the testing composites, and if so quantify the number of retrievals that return the query target or a near-duplicate; ideally, evaluate with a held-out gallery that does not contain the target images used for testing.
  4. [§III-B, Eqs. (3)-(5); Table III] The retrieval criterion is not validated independently. The thresholds epsilon_c=0.7 and epsilon_a=0.9 are set empirically with no sensitivity analysis, and there is no quantitative measure of retrieval precision or of whether 'illumination consistent' references actually improve harmonization beyond a random reference. Table III only varies epsilon_a in coarse bins; it does not report retrieval recall or precision. Please add a retrieval evaluation (e.g., precision@k against human-labeled content and illumination consistency) and an ablation with random or content-only references.
minor comments (5)
  1. [Abstract] There is a typo 'andetc' in the abstract that should be 'and etc.'
  2. [§IV-B and throughout] The benchmark name is inconsistent: 'RAHarmony4' appears as 'Raiharmony4' in Sec. I and as 'RAHarnomy4' in Sec. IV-B. Please standardize the spelling.
  3. [§IV-C] HDNet is cited as [10] in Sec. II-A but as [48] in Sec. IV-C, while reference [48] is listed as LEMART, not HDNet. Please unify the citations.
  4. [Table I] Table I reports no standard deviations for Raiha, while Table II does; please report mean and standard deviation over multiple runs for all tables.
  5. [§III-C(c)] The sentence 'Note that augmented references encompass the foreground region' is unclear: because the augmented references are random crops of the target image, they often contain the foreground object itself. This should be stated explicitly, since it is directly relevant to the training-contamination concern in the non-reference evaluation.

Circularity Check

1 steps flagged · score 6.0 of 10

RAHarmony4's retrieval-augmented benchmark is built from ground-truth target images and can retrieve the paired target as the reference, so the Table II gain partly reduces to copying from the answer image.

  1. fitted input called prediction [Sec. III-C(c) Data Augmentation; Sec. IV-A Datasets; Sec. IV-B Implementation and Evaluation Details; Table II]
    "Target images inherently have consistent illumination with composite images. ... we apply random cropping, flipping, and resizing operations to the target image to create diverse references. ... we collect target images from the training set of iHarmony4 to construct the RAHarmony4 dataset ... While a composite image has multiple references produced by our retrieval method, we randomly select a reference and calculate the metrics."

    The paper's desired-reference criterion in Eq. (1) is c_r ≈ c_f and i_r ≈ i_b. For any RAHarmony4 composite, its own ground-truth target image satisfies both conditions exactly: the target foreground is the harmonized version of the composite foreground, and the target background is pixel-identical to the composite background. Since the RAHarmony4 gallery is constructed from target images of the same iHarmony4 split and no exclusion of the paired target is stated, the retrieval scoring of Eqs. (3)-(5) ranks that paired target as the maximally suitable reference. The model is trained with target-derived references in Sec. III-C(c) and evaluated with references drawn from target-image galleries in Sec.

full rationale

The clearest construction-level circularity is in the RAHarmony4 retrieval-augmented evaluation: the gallery is built from ground-truth target images, and the paper does not state that a query's paired target is excluded, so the 'retrieved reference' can be the very image the network is asked to predict. This makes the Table II SOTA claim partly self-referential. The non-reference Table I claim is not itself circular at test time, because the network receives only the composite when evaluated; however, the training protocol in Sec. III-C(c) gives Raiha auxiliary input derived from the target image while baselines do not receive that channel, which is a legitimate comparison confound but not a reduction-by-definition. I found no load-bearing self-citation chain: references to DVT and other prior works are external, and the method is tested against standard iHarmony4 baselines. The score of 6 reflects that one of the two central SOTA claims, the retrieval-augmented one, is partially forced by the benchmark construction, while the non-reference result retains independent empirical content modulo the training-protocol concern.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim leans on three unproven premises: the Retinex-based equivalence between content+appearance matching and illumination consistency, the reliability of DVT features for harmonization-relevant semantics, and the appropriateness of a gallery built from ground-truth target images. The thresholds ϵc and ϵa are hand-set and not validated externally.

free parameters (2)
  • ϵc = 0.7
    Content similarity threshold in Eq. (3), empirically set to decide when a foreground patch and a gallery patch are semantically similar; not learned.
  • ϵa = 0.9
    Appearance similarity threshold in Eq. (5), empirically set to accept a reference as illumination-consistent via HSV histogram matching; hand-tuned.
assumptions (3)
  • domain assumption Retinex theory: appearance can be decomposed into reflectance and illumination components, and matching content plus appearance determines illumination consistency.
    Invoked in Sec. III-B to justify the retrieval criterion; ignores shadows, cast lighting, and sensor effects.
  • domain assumption DVT dense visual features provide reliable semantic correspondence for harmonization.
    Used in Eqs. (2), (3), and (8) without task-specific fine-tuning or validation of the correspondence quality.
  • domain assumption A gallery of ground-truth target images from the iHarmony4 training set is a valid reference source for evaluating retrieval-augmented harmonization.
    Assumes the availability of a curated set of correctly harmonized images; in real-world use such a gallery may not exist.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Retrieval Augmented Image Harmonization." pith.science (2026). https://pith.science/paper/H57YOZ4D

@misc{pith2026241213916,
  author       = {Pith},
  title        = {Pith review of: Retrieval Augmented Image Harmonization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H57YOZ4D}},
  note         = {Machine review of arXiv:2412.13916}
}
read the original abstract

When embedding objects (foreground) into images (background), considering the influence of photography conditions like illumination, it is usually necessary to perform image harmonization to make the foreground object coordinate with the background image in terms of brightness, color, and etc. Although existing image harmonization methods have made continuous efforts toward visually pleasing results, they are still plagued by two main issues. Firstly, the image harmonization becomes highly ill-posed when there are no contents similar to the foreground object in the background, making the harmonization results unreliable. Secondly, even when similar contents are available, the harmonization process is often interfered with by irrelevant areas, mainly attributed to an insufficient understanding of image contents and inaccurate attention. As a remedy, we present a retrieval-augmented image harmonization (Raiha) framework, which seeks proper reference images to reduce the ill-posedness and restricts the attention to better utilize the useful information. Specifically, an efficient retrieval method is designed to find reference images that contain similar objects as the foreground while the illumination is consistent with the background. For training the Raiha framework to effectively utilize the reference information, a data augmentation strategy is delicately designed by leveraging existing non-reference image harmonization datasets. Besides, the image content priors are introduced to ensure reasonable attention. With the presented Raiha framework, the image harmonization performance is greatly boosted under both non-reference and retrieval-augmented settings. The source code and pre-trained models will be publicly available.

Figures

Figures reproduced from arXiv: 2412.13916 by the authors.

Figure 1
Figure 1. When the background contains similar content to the foreground, it can provide internal guidance for adjusting foreground appearance. Conversely, image harmonization be￾comes ill-posed. Providing references that have similar content to the foreground can alleviate this issue. However, while the illumination between the reference and background is different, it fails to provide reliable guidance. Thus, when a referen… view at source ↗
Figure 2
Figure 2. (a) presents the framework of our proposed Raiha, which solves the ill-posed issues when similar content is unavailable in the background xb. Firstly, given a composite image x˜, we design a harmonization-oriented image retrieval method to produce proper a reference xr from the gallery G. Specifically, a foreground content retrieval is designed to construct the gallery Gsc, in which images contain similar content wi… view at source ↗
Figure 3
Figure 3. Qualitative comparison results on the RAHarmony4 dataset (the retrieval-augmented setting of the iHarmony dataset). Raiha produces more visually consistent results than other SOTA methods. sic [45], D-HT [46], CDTNet [7], Harmonizer [5], DCCF [4], PCTNet [47] and HDNet [48]. Note that all methods are trained and tested in the resolution of 256 × 256. Firstly, we conduct comparison experiments on the iHar￾mony4 datas… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 48 canonical work pages

  1. [1]

    Dovenet: Deep image harmonization via domain verification,

    W. Cong, J. Zhang, L. Niu, L. Liu, Z. Ling, W. Li, and L. Zhang, “Dovenet: Deep image harmonization via domain verification,” inCVPR, 2020, pp. 8391–8400. 1, 2, 5

    work page 2020

  2. [2]

    Bargainnet: Background-guided domain translation for image harmonization,

    W. Cong, L. Niu, J. Zhang, J. Liang, and L. Zhang, “Bargainnet: Background-guided domain translation for image harmonization,” in ICME, 2021, pp. 1–6. 1, 2, 5

    work page 2021

  3. [3]

    Region-aware adaptive instance normalization for image harmonization,

    J. Ling, H. Xue, L. Song, R. Xie, and X. Gu, “Region-aware adaptive instance normalization for image harmonization,” in CVPR, 2021, pp. 9361–9370. 1, 2, 5

    work page 2021

  4. [4]

    DCCF: deep comprehensible color filter learning framework for high-resolution image harmonization,

    B. Xue, S. Ran, Q. Chen, R. Jia, B. Zhao, and X. Tang, “DCCF: deep comprehensible color filter learning framework for high-resolution image harmonization,” in ECCV, 2022. 1, 2, 6

    work page 2022

  5. [5]

    Harmonizer: Learning to perform white-box image and video harmonization,

    Z. Ke, C. Sun, L. Zhu, K. Xu, and R. W. Lau, “Harmonizer: Learning to perform white-box image and video harmonization,” in ECCV, 2022, pp. 690–706. 1, 2, 6

    work page 2022

  6. [6]

    Spatial-separated curve rendering network for efficient and high-resolution image harmonization,

    J. Liang, X. Cun, and C.-M. Pun, “Spatial-separated curve rendering network for efficient and high-resolution image harmonization,” in ECCV, 2022. 1

    work page 2022

  7. [7]

    High- resolution image harmonization via collaborative dual transformations,

    W. Cong, X. Tao, L. Niu, J. Liang, X. Gao, Q. Sun, and L. Zhang, “High- resolution image harmonization via collaborative dual transformations,” in CVPR, 2022, pp. 18 470–18 479. 1, 2, 6

    work page 2022

  8. [8]

    Image Harmonization by Matching Regional References

    Z. Zhu, Z. Zhang, Z. Lin, R. Wu, Z. Chai, and C.-L. Guo, “Im- age harmonization by matching regional references,” arXiv preprint arXiv:2204.04715, 2022. 1, 2

    work page Pith review arXiv 2022

Show all 50 references
  1. [9]

    Scs-co: Self-consistent style contrastive learning for image harmonization,

    Y . Hang, B. Xia, W. Yang, and Q. Liao, “Scs-co: Self-consistent style contrastive learning for image harmonization,” in CVPR, 2022, pp. 19 710–19 719. 1, 2

    2022

  2. [10]

    Hierar- chical dynamic image harmonization,

    H. Chen, Z. Gu, Y . Li, J. Lan, C. Meng, W. Wang, and H. Li, “Hierar- chical dynamic image harmonization,” arXiv preprint arXiv:2211.08639,

    arXiv

  3. [11]

    Deep learning for instance retrieval: A survey,

    W. Chen, Y . Liu, W. Wang, E. M. Bakker, T. Georgiou, P. Fieguth, L. Liu, and M. S. Lew, “Deep learning for instance retrieval: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2022. 2, 3

    2022

  4. [12]

    Cnn features off-the-shelf: an astounding baseline for recognition,

    A. Sharif Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, “Cnn features off-the-shelf: an astounding baseline for recognition,” in Pro- ceedings of the IEEE conference on computer vision and pattern recognition workshops, 2014, pp. 806–813. 2, 3

    2014

  5. [13]

    Controlvideo: Training-free controllable text-to-video generation,

    Y . Zhang, Y . Wei, D. Jiang, X. ZHANG, W. Zuo, and Q. Tian, “Controlvideo: Training-free controllable text-to-video generation,” in The Twelfth International Conference on Learning Representations . 2

  6. [14]

    De- noising vision transformers,

    J. Yang, K. Z. Luo, J. Li, K. Q. Weinberger, Y . Tian, and Y . Wang, “De- noising vision transformers,” arXiv preprint arXiv:2401.02957, 2024. 2, 3, 4

    2024 arXiv

  7. [15]

    SSAT: A symmetric semantic-aware transformer network for makeup transfer and removal,

    Z. Sun, Y . Chen, and S. Xiong, “SSAT: A symmetric semantic-aware transformer network for makeup transfer and removal,” in AAAI, 2022, pp. 2325–2334. 2

    2022

  8. [16]

    Lemart: Label-efficient masked region transform for image harmonization,

    S. Liu, C. P. Huynh, C. Chen, M. Arap, and R. Hamid, “Lemart: Label-efficient masked region transform for image harmonization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18 290–18 299. 2

    2023

  9. [17]

    Neural codes for image retrieval,

    A. Babenko, A. Slesarev, A. Chigorin, and V . Lempitsky, “Neural codes for image retrieval,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. Springer, 2014, pp. 584–599. 3

    2014

  10. [18]

    A baseline for visual instance retrieval with deep convolutional networks,

    S. Ali, J. Sullivan, A. Maki, and S. Carlsson, “A baseline for visual instance retrieval with deep convolutional networks,” in Proceedings of International Conference on Learning Representations , 2015. 3

    2015

  11. [19]

    Embedding based on function approxima- tion for large scale image search,

    T.-T. Do and N.-M. Cheung, “Embedding based on function approxima- tion for large scale image search,” IEEE transactions on pattern analysis and machine intelligence , vol. 40, no. 3, pp. 626–638, 2017. 3

    2017

  12. [20]

    Spatial pyramid deep hashing for large-scale image retrieval,

    W. Zhao, H. Luo, J. Peng, and J. Fan, “Spatial pyramid deep hashing for large-scale image retrieval,” Neurocomputing, vol. 243, pp. 166–173,

  13. [21]

    Deep image retrieval: Learning global representations for image search,

    A. Gordo, J. Almaz ´an, J. Revaud, and D. Larlus, “Deep image retrieval: Learning global representations for image search,” in Computer Vision– ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14 . Springer, 2016, pp. 2...

    2016

  14. [22]

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

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” Advances in neural information processing systems , vol. 28, 2015. 3

    2015

  15. [23]

    Class-weighted convolutional features for visual instance search,

    A. Jimenez, J. M. Alvarez, and X. Giro-i Nieto, “Class-weighted convolutional features for visual instance search,” arXiv preprint arXiv:1707.02581, 2017. 3

    2017 arXiv

  16. [24]

    Video google: A text retrieval approach to object matching in videos,

    Sivic and Zisserman, “Video google: A text retrieval approach to object matching in videos,” in Proceedings ninth IEEE international conference on computer vision . IEEE, 2003, pp. 1470–1477. 3

    2003

  17. [25]

    Aggregating local de- scriptors into a compact image representation,

    H. J ´egou, M. Douze, C. Schmid, and P. P ´erez, “Aggregating local de- scriptors into a compact image representation,” in 2010 IEEE computer society conference on computer vision and pattern recognition . IEEE, 2010, pp. 3304–3311. 3

    2010

  18. [26]

    Fisher kernels on visual vocabularies for image categorization,

    F. Perronnin and C. Dance, “Fisher kernels on visual vocabularies for image categorization,” in 2007 IEEE conference on computer vision and pattern recognition. IEEE, 2007, pp. 1–8. 3

    2007

  19. [27]

    Local convolutional features with unsupervised training for image retrieval,

    M. Paulin, M. Douze, Z. Harchaoui, J. Mairal, F. Perronin, and C. Schmid, “Local convolutional features with unsupervised training for image retrieval,” in Proceedings of the IEEE international conference on computer vision , 2015, pp. 91–99. 3

    2015

  20. [28]

    Image search with selective match kernels: aggregation across single and multiple images,

    G. Tolias, Y . Avrithis, and H. J´egou, “Image search with selective match kernels: aggregation across single and multiple images,” International Journal of Computer Vision , vol. 116, pp. 247–261, 2016. 3

    2016

  21. [29]

    Unifying deep local and global features for image search,

    B. Cao, A. Araujo, and J. Sim, “Unifying deep local and global features for image search,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX

    2020

  22. [30]

    Retrieval oriented deep feature learning with complementary supervision mining,

    Y . Lv, W. Zhou, Q. Tian, S. Sun, and H. Li, “Retrieval oriented deep feature learning with complementary supervision mining,” IEEE Transactions on Image Processing, vol. 27, no. 10, pp. 4945–4957, 2018. 3

    2018

  23. [31]

    Springer, 2020, pp. 726–743. 3

    2020

  24. [32]

    Learning with average precision: Training image retrieval with a listwise loss,

    J. Revaud, J. Almaz ´an, R. S. Rezende, and C. R. d. Souza, “Learning with average precision: Training image retrieval with a listwise loss,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 5107–5116. 3

    2019

  25. [33]

    Deep clustering for unsupervised learning of visual features,

    M. Caron, P. Bojanowski, A. Joulin, and M. Douze, “Deep clustering for unsupervised learning of visual features,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 132–149. 3

    2018

  26. [34]

    Fine-tuning cnn image retrieval with no human annotation,

    F. Radenovi ´c, G. Tolias, and O. Chum, “Fine-tuning cnn image retrieval with no human annotation,” IEEE transactions on pattern analysis and machine intelligence, vol. 41, no. 7, pp. 1655–1668, 2018. 3

    2018

  27. [35]

    Deephash for image instance retrieval: Getting regulariza- tion, depth and fine-tuning right,

    J. Lin, O. Morere, A. Veillard, L.-Y . Duan, H. Goh, and V . Chan- drasekhar, “Deephash for image instance retrieval: Getting regulariza- tion, depth and fine-tuning right,” in Proceedings of the 2017 ACM on International Conference on Multimedia Retrieval , 2017, pp. 133–141. 3

    2017

  28. [36]

    A two-stage triplet network training framework for image retrieval,

    W. Min, S. Mei, Z. Li, and S. Jiang, “A two-stage triplet network training framework for image retrieval,” IEEE Transactions on Multimedia , vol. 22, no. 12, pp. 3128–3138, 2020. 3

    2020

  29. [37]

    Multiple saliency and channel sensitivity network for aggregated convolutional feature,

    X. Xiang, Z. Wang, Z. Zhao, and F. Su, “Multiple saliency and channel sensitivity network for aggregated convolutional feature,” inProceedings of the AAAI Conference on Artificial Intelligence , vol. 33, no. 01, 2019, pp. 9013–9020. 3

    2019

  30. [38]

    Fast spectral ranking for similarity search,

    A. Iscen, Y . Avrithis, G. Tolias, T. Furon, and O. Chum, “Fast spectral ranking for similarity search,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 7632–7641. 3

    2018

  31. [39]

    Mining on manifolds: Metric learning without labels,

    A. Iscen, G. Tolias, Y . Avrithis, and O. Chum, “Mining on manifolds: Metric learning without labels,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 7642–7651. 3

    2018

  32. [40]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763. 3

    2021

  33. [41]

    Deep convolutional image retrieval: A general framework,

    M. Tzelepi and A. Tefas, “Deep convolutional image retrieval: A general framework,” Signal Processing: Image Communication, vol. 63, pp. 30– 43, 2018. 3

    2018

  34. [42]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in MICCAI, vol. 9351, 2015, pp. 234–241. 4

    2015

  35. [43]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, R. Howes, P.-Y . Huang, H. Xu, V . Sharma, S.-W. Li, W. Galuba, M. Rabbat, M. Assran, N. Ballas, G. Synnaeve, I. Misra, H. Jegou, J. Mairal, P. Labatu...

    2023

  36. [44]

    Foreground-aware semantic representations for image harmonization,

    K. Sofiiuk, P. Popenova, and A. Konushin, “Foreground-aware semantic representations for image harmonization,” in WACV, 2021, pp. 1620–

    2021

  37. [45]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al. , “Pytorch: An imperative style, high-performance deep learning library,” NeurIPS, vol. 32, 2019. 5

    2019

  38. [46]

    Image harmonization with transformer,

    Z. Guo, D. Guo, H. Zheng, Z. Gu, B. Zheng, and J. Dong, “Image harmonization with transformer,” in ICCV, 2021, pp. 14 850–14 859. 6

    2021

  39. [47]

    Intrinsic image harmonization,

    Z. Guo, H. Zheng, Y . Jiang, Z. Gu, and B. Zheng, “Intrinsic image harmonization,” in CVPR, 2021, pp. 16 367–16 376. 6

    2021

  40. [48]

    Lemart: Label-efficient masked region transform for image harmonization,

    S. Liu, C. P. Huynh, C. Chen, M. Arap, and R. Hamid, “Lemart: Label-efficient masked region transform for image harmonization,”arXiv preprint arXiv:2304.13166, 2023. 6

    2023 arXiv

  41. [49]

    Pct-net: Full res- olution image harmonization using pixel-wise color transformations,

    J. J. A. Guerreiro, M. Nakazawa, and B. Stenger, “Pct-net: Full res- olution image harmonization using pixel-wise color transformations,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 5917–5926. 6

    2023

  42. [2017]

    3 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 8

    2021

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.