Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Improving Zero-Shot Object-Level Change Detection by Incorporating Visual Correspondence

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that training a change detector with correspondence supervision improves its accuracy, and that a homography-based alignment plus Hungarian matching at test time removes false positives and, for the first time, links…

desk verdict Useful idea and a new benchmark, but the zero-shot numbers are selected on the test sets, so the headline margins are optimistic. read the letter →

arxiv 2501.05555 v2 pith:3Q4QJLKT submitted 2025-01-09 cs.CV cs.AI

Hung Huy Nguyen , Pooyan Rahmanzadehgervi , Long Mai , Anh Totti Nguyen This is my paper · ORCID
classification cs.CVcs.AI
keywords changedetectionobject-levelvisualcorrespondencecontrastivelosshomographyalignmentHungarianalgorithmzero-shotgeneralizationOpenImages-Inpainted
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

The paper tries to establish that visual correspondence is more than a useful output of change detection: supervising a detector to recognize which boxes across two images refer to the same changed object makes the detector itself more accurate, and a test-time alignment step built on that same idea removes false positives. If the claim holds, image-inspection and surveillance systems would get a list of changed objects together with explicit links between the two views, and they would raise fewer alarms on image pairs with no change. The reported experiments cover five datasets, four of them unseen during training, with the fine-tuned model outperforming the prior detector on every benchmark and at every post-processing stage. The paper also introduces a large single-change dataset and an F1-style metric for scoring predicted correspondences.

What carries the argument

The mechanism that carries the argument is bidirectional visual correspondence. A contrastive matching loss supervises the network with binary labels of whether two boxes are the same change, and an alignment stage at inference filters detections by projecting boxes across a homography estimated from point matches; the Hungarian algorithm then completes the pairing. Named components: the contrastive matching loss, the homography-based alignment stage, and Hungarian bipartite matching over pooled encoder embeddings.

What would settle it

An evaluation on image pairs with large 3D parallax, non-planar scenes, or repeated textures, comparing recall with the alignment stage disabled, would show whether the homography assumption discards true positives; if a detector trained with the same contrastive loss but without homography alignment matches the reported accuracy, the alignment premise is not the source of the gains.

Watch

Extended reading notes

Core claim

The central claim is that change correspondences should be used twice: as a training signal and as a test-time filter. During fine-tuning of a pretrained change detector, a contrastive matching loss is computed on embeddings of ground-truth matched boxes, teaching the network to distinguish true matches from spurious ones; this improves detection mAP on all five benchmarks and also improves the F1 score of the predicted correspondences. At inference, the pipeline has three stages: the detector proposes boxes, a homography estimated from learned point matches (SuperGlue plus RANSAC) projects each proposed box from one image into the other and discards any box whose projection overlaps no box in the other image, and the Hungarian algorithm pairs the surviving boxes using cosine distance between pooled encoder embeddings. The paper further claims to be the first to output these change-to-change correspondences, and it introduces OpenImages-Inpainted, roughly 1.3 million image pairs with exactly one inpainted-object change, to support training and evaluation.

Load-bearing premise

The alignment stage assumes the two images are related by a single homography and that every genuine change, including a removed object, leaves a box footprint that projects into the other image with some overlap; when viewpoint or scene geometry violates that, true changes are filtered out together with false positives.

Editorial extensions

If this is right

  • Fine-tuning with the contrastive matching loss improves change-detection mAP on all five benchmarks, including four unseen zero-shot datasets, relative to the baseline detector at every post-processing stage.
  • The three-stage post-processing pipeline (confidence threshold, homography alignment, Hungarian matching) reduces the average number of predicted boxes per image on no-change pairs, lowering false-positive alarms.
  • Change detection systems can output explicit correspondences between detected objects in the two views, not just independent box lists.
  • The proposed F1-based correspondence metric allows detectors to be compared on matching quality, not only localization, for the first time.
  • The OpenImages-Inpainted dataset provides a large, single-change training and evaluation resource with fewer view-transformation artifacts than prior synthetic change datasets.

Reading between the lines

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

  • The same correspondence-supervision idea could carry over to change segmentation or panoptic change detection, where per-pixel correspondence masks would replace box-level matching and could be evaluated with a pixel-level variant of the proposed F1 metric.
  • Because the homography alignment is the stated fragile point, replacing the explicit RANSAC homography with a learned geometric verification or a differentiable matching layer could extend the method to non-planar scenes and large viewpoint changes.
  • The reported false-positive reduction suggests the network learns an implicit 'no-change' notion from the contrastive loss; a direct ablation that removes no-change pairs from fine-tuning would test that explanation.
  • The F1 correspondence metric might transfer to change-captioning evaluation, scoring whether a caption's described change correctly points to the matching boxes across the image pair.
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 / 5 minor

Summary. The paper proposes a method for object-level change detection that fine-tunes the CYWS change detector with a contrastive matching loss computed from change-correspondence labels, and then applies a test-time post-processing pipeline: confidence thresholding, homography-based alignment of predicted boxes using SuperGlue and RANSAC, and Hungarian-algorithm matching between the two images. The paper claims state-of-the-art mAP on COCO-Inpainted and four zero-shot benchmarks, improved correspondence F1, consistent false-positive evaluation on no-change pairs, and the first change-correspondence prediction framework. It also introduces OpenImages-Inpainted, a large-scale dataset of about 1.3M pairs with exactly one change per pair.

Significance. If the claims are supported, the paper makes several useful contributions: a novel training signal (contrastive matching loss) that improves both detection and correspondence, a post-processing stage that reduces false positives on no-change inputs, a new evaluation metric for change correspondences, and a large new dataset. The code and data are promised on GitHub, which helps reproducibility. However, the zero-shot generalization claim is not established by the current experimental protocol, because multiple design choices and hyperparameters are selected using the same five benchmarks on which zero-shot performance is reported. The in-distribution gains are more credible, but the central 'zero-shot' claim needs a validation-based selection protocol before it can be accepted.

major comments (3)
  1. [Section 4.1 (Table 1), Section 4.2 (Table 2), Appendices C-D (Tables A3-A5)] The zero-shot claim is compromised by selection on the test benchmarks. The manuscript chooses the embedding assignment method (Table 1), encoder versus decoder features (Table 2), learning rate (Table A3), number of epochs (Table A4), and decoder depth (Table A5) by comparing mAP/F1 on the same five benchmarks on which the final zero-shot results are reported. For example, Appendix C states that 500 epochs were rejected because they reduced zero-shot accuracy on the four zero-shot datasets, and Appendix D rejects a deeper decoder based on test mAP. The reported margins (e.g., +1.05 to +9.04 in mAP) are therefore test-selected upper bounds, and the paper does not establish that the benefit of the contrastive loss transfers to unseen domains independently of this selection. A validation-based selection protocol (e.g., using a held-out split of COCO-Inpainted) is required before the zero-shot claim can be accepted.
  2. [Section 5.1 and Table 5] The stated operating-point fairness criterion is not substantiated. Section 5.1 says a detection threshold of 0.25 was chosen to keep the average number of predicted boxes per image on no-change pairs below 0.01, but Table 5, the only no-change quantitative evidence, reports values at threshold 0.1 and shows full-pipeline averages between 0.06 and 1.156 boxes per image; no row at threshold 0.25 is reported and no value below 0.01 appears. The authors should either report the no-change box counts at the chosen threshold or revise the stated fairness criterion.
  3. [Section 3.3 and Table 4] The sentence in Section 3.3 claiming that the alignment stage combined with confidence thresholding 'yields improved mAP' is contradicted by Table 4. For both models, adding alignment and then Hungarian matching monotonically decreases mAP at threshold 0.25 (e.g., CYWS on COCO-Inpainted drops from 51.25 to 47.83 to 44.48 in rows b, c, and d). The authors should correct this statement to refer to false-positive reduction in no-change cases (Table 5) rather than mAP, or provide the specific mAP result they intended.
minor comments (5)
  1. [Tables 1-7 and Tables A1-A5] The column headers for the five benchmarks appear as blank spaces in the submitted PDF; please ensure the dataset names render correctly in the final version.
  2. [Section 2.3] The use of 'optimal detection threshold' in mAP evaluation is non-standard, because mAP normally integrates over confidence thresholds; clarify that predictions are first filtered by confidence and mAP is computed on the remaining boxes.
  3. [Equation (2) and Section 3.4] The loss weights alpha and beta in Equation (2) are fixed at 3 and 2, but no sensitivity analysis is reported for them; adding such an analysis would strengthen the loss ablation in Section 4.3.
  4. [Section 2.3] The TP/FP/FN definitions for correspondence evaluation should clarify the matching protocol when multiple predicted pairs have nonzero IoU with the same ground-truth pair, beyond the 'highest IoU' rule.
  5. [Section 7 and Section 3.2] The limitations paragraph concedes that point-estimation accuracy declines under significant distortions or detailed textures; quantifying this failure rate on the five benchmarks would help readers judge when the alignment stage can be trusted in practice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the contrastive loss, alignment, and Hungarian matching are supervised and evaluated on held-out ground truth; test-set hyperparameter selection is a validity concern, not circular reasoning.

full rationale

No equation-level circularity is present in the claimed derivation chain. The contrastive matching loss (Eq. 2) is a supervised objective whose positive and negative pairs are built by comparing Hungarian matches of ground-truth boxes against ground-truth correspondence labels (Fig. 5); at test time, embeddings, Hungarian matching, and the F1 correspondence metric are all evaluated against held-out ground-truth boxes, so the reported gains are not equal to the training labels by construction. Stage 2 alignment is an independent geometric filter: the homography is estimated from SuperGlue matches with RANSAC, and candidate boxes are kept only when their projected box has IoU > 0 with a detected box in the other image; this does not assume the correspondence output it claims to produce. The Hungarian matching stage minimizes a cosine-distance cost, and the same post-processing is applied to the CYWS baseline, so the comparison is not an identity. The only self-referential element is methodological: several design choices (embedding assignment, encoder versus decoder features, epoch count, learning rate, decoder depth) are selected using mAP/F1 on the zero-shot benchmarks themselves (Sec. 4.1-4.2, App. B-D), which is a test-set selection concern and a threat to the generality of the reported zero-shot margins, but it is not a case of a prediction being equivalent to its input by definition or a fitted parameter being relabeled as a prediction. The paper also acknowledges the fragility of the homography-based alignment in Sec. 7, which is a limitation rather than a circular step. Citation [23] by one of the authors is used only for the standard cosine ground-distance formula and is not load-bearing. Therefore, the circularity score is 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

Most of the empirical claim rests on supervised finetuning on COCO-Inpainted plus post-processing; the largest unstated premise is the homography alignment assumption in Stage 2. The ledger also records the key hand-set hyperparameters and the evaluation-driven model selection that weakens the zero-shot claim.

free parameters (7)
  • alpha (DETR loss weight in Eq. (2)) = 3
    Hand-set; no sensitivity analysis beyond the on/off ablation in Table 3.
  • beta (contrastive matching loss weight in Eq. (2)) = 2
    Hand-set; sensitivity not reported.
  • Detection confidence threshold = 0.25
    Chosen in Sec. 5.1 to keep average no-change boxes below 0.01 per image, i.e., selected on no-change test data.
  • IoU threshold for matching TP/FP = 0.5
    Standard but hand-set criterion in Sec. 2.3 for the new correspondence F1 metric.
  • Object relative-area range for OpenImages-Inpainted = 0.01 to 0.04
    Dataset construction filter in Sec. 2.2; narrows objects to a difficulty band and affects benchmark results.
  • Learning rate = 0.0001
    Chosen over 0.0005 because 0.0005 reduced mAP on the benchmarks (App. C, Table A3), i.e., selected on zero-shot test data.
  • Number of training epochs = 200
    Chosen over 500 because 500 reduced zero-shot accuracy on four benchmarks (App. C, Table A4).
assumptions (5)
  • domain assumption Two images of the same scene can be aligned by a single homography (affine transformation) estimated from SuperGlue point matches and RANSAC.
    This premise underlies Stage 2 alignment (Sec. 3.2, Fig. 4); the paper's Limitations section concedes point-estimation accuracy declines with significant distortions or detailed textures.
  • domain assumption Every true change has a projectable counterpart box in the other image with IoU > 0.
    Stage 2 discards predicted boxes whose projected counterpart overlaps no box in the other image (Sec. 3.2); true changes that fail this test are removed from the output.
  • domain assumption Ground-truth change and correspondence labels in COCO-Inpainted are complete and accurate.
    These labels train the contrastive matching loss (Sec. 3.3) and define the F1 metric (Sec. 2.3); no label-quality analysis is provided.
  • ad hoc to paper Selecting hyperparameters with mAP/F1 on the zero-shot benchmarks is a valid way to measure zero-shot generalization.
    The paper chooses embedding method, feature source, learning rate, number of epochs, loss weights, and detection threshold using all five benchmarks, including the zero-shot test sets (Sec. 4.1, 4.2, App. C); this assumption is load-bearing for the claimed zero-shot margins.
  • ad hoc to paper An average of fewer than 0.01 predicted boxes per image on no-change pairs is the correct operating-point fairness criterion.
    Used to set detection threshold 0.25 in Sec. 5.1; it is a reasonable deployment criterion but is chosen without a stated protocol and is not reported for the proposed model at 0.25.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Zero-Shot Object-Level Change Detection by Incorporating Visual Correspondence." pith.science (2026). https://pith.science/paper/3Q4QJLKT

@misc{pith2026250105555,
  author       = {Pith},
  title        = {Pith review of: Improving Zero-Shot Object-Level Change Detection by Incorporating Visual Correspondence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3Q4QJLKT}},
  note         = {Machine review of arXiv:2501.05555}
}
read the original abstract

Detecting object-level changes between two images across possibly different views is a core task in many applications that involve visual inspection or camera surveillance. Existing change-detection approaches suffer from three major limitations: (1) lack of evaluation on image pairs that contain no changes, leading to unreported false positive rates; (2) lack of correspondences (i.e., localizing the regions before and after a change); and (3) poor zero-shot generalization across different domains. To address these issues, we introduce a novel method that leverages change correspondences (a) during training to improve change detection accuracy, and (b) at test time, to minimize false positives. That is, we harness the supervision labels of where an object is added or removed to supervise change detectors, improving their accuracy over previous work by a large margin. Our work is also the first to predict correspondences between pairs of detected changes using estimated homography and the Hungarian algorithm. Our model demonstrates superior performance over existing methods, achieving state-of-the-art results in change detection and change correspondence accuracy across both in-distribution and zero-shot benchmarks.

Figures

Figures reproduced from arXiv: 2501.05555 by the authors.

Figure 1
Figure 1. At an optimal confidence threshold, CYWS [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Example predicted changes. In , our method detects a removal of an elephant (a) despite viewpoint and color differences between two images. A tiny, white object (b) removal is detected despite viewpoint differences in . In , a blue truck moving from one location to another is correctly detected as two changes: a removal and an addition. Colored, solid-bordered boxes, e.g., (□, □), (□, □), (□, □), show predicted pair… view at source ↗
Figure 3
Figure 3. Our approach of detecting changes and predicting their correspondence. Our approach comprises three major stages. The first is the change detector, which we employ from the CYWS paper to identify changes between two images. The second is the alignment step, where an ideal detection threshold is established before forwarding anticipated boxes to the alignment process, aiding in the removal of false positive predicted… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Alignment Overview. (Fig. 4a) illustrates the process for estimating the homography matrix. In (Fig. 4b), dashed □ represent the corresponding box in the right image of □ in the left image. Similarly for □ box. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Contrastive matching loss. Contrastive matching loss trains the model to distinguish between positive and negative pairs of image patches. Each ground-truth bounding box is assigned an embedding from the last layer of encoder, and the Hungarian algorithm is employed to…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 29 canonical work pages

  1. [1]

    Street-view change detec- tion with deconvolutional networks

    Pablo F Alcantarilla, Simon Stent, German Ros, Roberto Arroyo, and Riccardo Gherardi. Street-view change detec- tion with deconvolutional networks. Autonomous Robots , 42:1301–1322, 2018. 8

    work page 2018

  2. [2]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European confer- ence on computer vision, pages 213–229. Springer, 2020. 2, 5

    work page 2020

  3. [3]

    Changes to Captions: An Attentive Network for Remote Sensing Change Captioning

    Shizhen Chang and Pedram Ghamisi. Changes to captions: An attentive network for remote sensing change captioning. arXiv preprint arXiv:2304.01091, 2023. 8

    work page Pith review arXiv 2023

  4. [4]

    Changemamba: Remote sensing change detection with spatio-temporal state space model

    Hongruixuan Chen, Jian Song, Chengxi Han, Junshi Xia, and Naoto Yokoya. Changemamba: Remote sensing change detection with spatio-temporal state space model. arXiv preprint arXiv:2404.03425, 2024. 8

    arXiv 2024

  5. [5]

    Joint spatio-temporal modeling for semantic change detection in remote sensing images

    Lei Ding, Jing Zhang, Haitao Guo, Kai Zhang, Bing Liu, and Lorenzo Bruzzone. Joint spatio-temporal modeling for semantic change detection in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 8

    work page 2024

  6. [6]

    Deepatrophy: Teaching a neural network to detect pro- gressive changes in longitudinal mri of the hippocampal re- gion in alzheimer’s disease

    Mengjin Dong, Long Xie, Sandhitsu R Das, Jiancong Wang, Laura EM Wisse, Robin DeFlores, David A Wolk, Paul A Yushkevich, Alzheimer’s Disease Neuroimaging Initiative, et al. Deepatrophy: Teaching a neural network to detect pro- gressive changes in longitudinal mri of the hippocampal re- gion in alzheimer’s disease. Neuroimage, 243:118514, 2021. 1

    work page 2021

  7. [7]

    Centernet: Keypoint triplets for object detection

    Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qing- ming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 6569–6578,

  8. [8]

    Neural naturalist: generat- ing fine-grained image comparisons

    Maxwell Forbes, Christine Kaeser-Chen, Piyush Sharma, and Serge Belongie. Neural naturalist: generat- ing fine-grained image comparisons. arXiv preprint arXiv:1909.04101, 2019. 8

    arXiv 1909

Show all 41 references
  1. [9]

    Superpixel-based difference representation learning for change detection in multispectral remote sensing im- ages

    Maoguo Gong, Tao Zhan, Puzhao Zhang, and Qiguang Miao. Superpixel-based difference representation learning for change detection in multispectral remote sensing im- ages. IEEE Transactions on Geoscience and Remote sensing, 55(5):2658–2673, 2017. 8

    2017

  2. [10]

    Clip4idc: Clip for image difference captioning

    Zixin Guo, Tzu-Jui Julius Wang, and Jorma Laaksonen. Clip4idc: Clip for image difference captioning. arXiv preprint arXiv:2206.00629, 2022. 8

    2022 arXiv

  3. [11]

    Image change cap- tioning by learning from an auxiliary task

    Mehrdad Hosseinzadeh and Yang Wang. Image change cap- tioning by learning from an auxiliary task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2725–2734, 2021. 8

    2021

  4. [12]

    Image difference captioning with instance-level fine-grained feature represen- tation

    Qingbao Huang, Yu Liang, Jielong Wei, Yi Cai, Hanyu Liang, Ho-fung Leung, and Qing Li. Image difference captioning with instance-level fine-grained feature represen- tation. IEEE transactions on multimedia , 24:2004–2017,

    2004

  5. [13]

    Learning to describe differences between pairs of similar images

    Harsh Jhamtani and Taylor Berg-Kirkpatrick. Learning to describe differences between pairs of similar images. arXiv preprint arXiv:1808.10584, 2018. 1, 2, 8

    2018 arXiv

  6. [14]

    Img-diff: Contrastive data synthesis for multimodal large language models

    Qirui Jiao, Daoyuan Chen, Yilun Huang, Yaliang Li, and Ying Shen. Img-diff: Contrastive data synthesis for multimodal large language models. arXiv preprint arXiv:2408.04594, 2024. 2

    2024 arXiv

  7. [15]

    Agnostic change captioning with cycle consistency

    Hoeseong Kim, Jongseok Kim, Hyungseok Lee, Hyunsung Park, and Gunhee Kim. Agnostic change captioning with cycle consistency. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 2095–2104,

  8. [16]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  9. [17]

    The hungarian method for the assignment problem

    Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly , 2(1-2):83–97,

  10. [18]

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Ste- fan Popov, Matteo Malloci, Tom Duerig, and Vittorio Ferrari. The open images dataset V4: unified image classification, object detection, and visual relationshi...

    2018 arXiv

  11. [19]

    Stade-cdnet: Spatial– temporal attention with difference enhancement-based net- work for remote sensing image change detection

    Zhi Li, Siying Cao, Jiakun Deng, Fengyi Wu, Ruilan Wang, Junhai Luo, and Zhenming Peng. Stade-cdnet: Spatial– temporal attention with difference enhancement-based net- work for remote sensing image change detection. IEEE Transactions on Geoscience and Remote Sensing , 62:1–17,

  12. [20]

    Padilla, S

    R. Padilla, S. L. Netto, and E. A. B. da Silva. A survey on performance metrics for object-detection algorithms. In 2020 International Conference on Systems, Signals and Im- age Processing (IWSSIP), pages 237–242, 2020. 3

    2020

  13. [21]

    Robust change captioning

    Dong Huk Park, Trevor Darrell, and Anna Rohrbach. Robust change captioning. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 4624–4633,

  14. [22]

    A review of the au- tomated detection of change in serial imaging studies of the brain

    Julia Patriarche and Bradley Erickson. A review of the au- tomated detection of change in serial imaging studies of the brain. Journal of digital imaging, 17:158–174, 2004. 1

    2004

  15. [23]

    Deepface-emd: Re-ranking using patch-wise earth mover’s distance improves out-of- distribution face identification

    Hai Phan and Anh Nguyen. Deepface-emd: Re-ranking using patch-wise earth mover’s distance improves out-of- distribution face identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20259–20269, 2022. 5

    2022

  16. [24]

    Concurrent spatial and channel ‘squeeze & excitation’in fully convolutional networks

    Abhijit Guha Roy, Nassir Navab, and Christian Wachinger. Concurrent spatial and channel ‘squeeze & excitation’in fully convolutional networks. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st In- ternational Conference, Granada, Spain, September ...

    2018

  17. [25]

    The change you want to see

    Ragav Sachdeva and Andrew Zisserman. The change you want to see. In Proceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 3993–4002,

  18. [26]

    The change you want to see (now in 3d)

    Ragav Sachdeva and Andrew Zisserman. The change you want to see (now in 3d). In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2060– 2069, 2023. 1, 2, 3, 6, 8 9

    2023

  19. [27]

    Unsupervised deep change vector analysis for multiple- change detection in vhr images

    Sudipan Saha, Francesca Bovolo, and Lorenzo Bruzzone. Unsupervised deep change vector analysis for multiple- change detection in vhr images. IEEE Transactions on Geo- science and Remote Sensing, 57(6):3677–3693, 2019. 8

    2019

  20. [28]

    Change detection from a street image pair using cnn features and superpixel segmen- tation

    Ken Sakurada and Takayuki Okatani. Change detection from a street image pair using cnn features and superpixel segmen- tation. In British Machine Vision Conference, 2015. 8

    2015

  21. [29]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 4

    2020

  22. [30]

    Finding it at another side: A viewpoint-adapted match- ing encoder for change captioning

    Xiangxi Shi, Xu Yang, Jiuxiang Gu, Shafiq Joty, and Jianfei Cai. Finding it at another side: A viewpoint-adapted match- ing encoder for change captioning. In Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, Au- gust 23–28, 2020, Proceedings, Part XIV 16, page...

    2020

  23. [31]

    Detecting change for multi-view, long-term surface inspection

    Simon Stent, Riccardo Gherardi, Bj ¨orn Stenger, and Roberto Cipolla. Detecting change for multi-view, long-term surface inspection. In BMVC, pages 127–1, 2015. 1

    2015

  24. [32]

    The stvchrono dataset: Towards continuous change recognition in time

    Yanjun Sun, Yue Qiu, Mariia Khan, Fumiya Matsuzawa, and Kenji Iwata. The stvchrono dataset: Towards continuous change recognition in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14111–14120, 2024. 8

    2024

  25. [33]

    Resolution-robust large mask inpainting with fourier convolutions

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE conferenc...

    2021

  26. [34]

    Expressing visual relationships via language

    Hao Tan, Franck Dernoncourt, Zhe Lin, Trung Bui, and Mo- hit Bansal. Expressing visual relationships via language. In Anna Korhonen, David Traum, and Llu ´ıs M`arquez, editors, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1873–...

    2019

  27. [35]

    Adaptive feature weighted fusion nested u-net with discrete wavelet transform for change detection of high-resolution remote sensing images

    Congcong Wang, Wenbin Sun, Deqin Fan, Xiaoding Liu, and Zhi Zhang. Adaptive feature weighted fusion nested u-net with discrete wavelet transform for change detection of high-resolution remote sensing images. Remote Sensing, 13(24):4971, 2021. 8

    2021

  28. [36]

    Spot the differ- ence by object detection

    Junhui Wu, Yun Ye, Yu Chen, and Zhi Weng. Spot the differ- ence by object detection. arXiv preprint arXiv:1801.01051,

    arXiv

  29. [37]

    Change capsule network for optical remote sensing image change detection

    Quanfu Xu, Keming Chen, Guangyao Zhou, and Xian Sun. Change capsule network for optical remote sensing image change detection. Remote Sensing, 13(14):2646, 2021. 1, 8

    2021

  30. [38]

    Deep siamese networks based change detection with remote sensing images

    Le Yang, Yiming Chen, Shiji Song, Fan Li, and Gao Huang. Deep siamese networks based change detection with remote sensing images. Remote Sensing, 13(17):3394, 2021. 8

    2021

  31. [39]

    Image difference captioning with pre-training and contrastive learning

    Linli Yao, Weiying Wang, and Qin Jin. Image difference captioning with pre-training and contrastive learning. InPro- ceedings of the AAAI Conference on Artificial Intelligence , volume 36, pages 3108–3116, 2022. 8

    2022

  32. [40]

    Segment any change

    Zhuo Zheng, Yanfei Zhong, Liangpei Zhang, and Ste- fano Ermon. Segment any change. arXiv preprint arXiv:2402.01188, 2024. 1, 2, 8 10 Appendix for: Improving Zero-Shot Object-Level Change Detection by Incorporating Visual Correspondence A. Upper bound accuracy of correspondence...

    2024 arXiv

  33. [2023]

    1, 2, 3, 5, 6, 7, 8, 13

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.