Pith. sign in

REVIEW 3 major objections 6 minor 34 references

This paper claims that image manipulation localization is improved by fusing monocular reconstructed depth and surface normals into RGB features only where they are judged reliable, and that this selective 3D evidence generalizes better tha

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

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

T0 review · deepseek-v4-flash

2026-08-01 16:14 UTC pith:O4NXH5YU

load-bearing objection GFrame makes a plausible, incremental case for reliability-gated monocular geometry in IML, but the ablation design leaves the central 'selective fusion' claim under-tested; worth expert review, not a desk reject. the 3 major comments →

arxiv 2607.18040 v1 pith:O4NXH5YU submitted 2026-07-20 cs.CV

When 2D Cues Fail: Improving Image Manipulation Localization with Reliable 3D Geometry

Guofeng Yu , Zhiqing Guo , Dan Ma , Gaobo Yang This is my paper
classification cs.CV
keywords image manipulation localizationmonocular 3D reconstructiondepth and surface normalsreliability-aware fusioncross-scale attentionmultimedia forensicsout-of-distribution generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

GFrame, the proposed framework, uses monocular depth and surface normals as auxiliary cues for manipulation localization but never trusts them uniformly. Its central principle—reconstructed geometry should be exploited selectively—is realized by a reliability gate that admits geometric features into the RGB stream only where three cheap confidence proxies (depth smoothness, normal unit-norm fidelity, depth-normal agreement) indicate trustworthy reconstruction. Under a matched 20-epoch training budget, the paper reports an average F1@0.5 of 0.592 on in-distribution benchmarks and 0.490 on out-of-distribution benchmarks, beating all compared appearance-driven baselines; the advantage persists under noise, blur, JPEG recompression, and social-media processing. A sympathetic reader would take this to establish that properly gated 3D geometry is a transferable complement to 2D forensics when manipulations are visually well blended.

Core claim

On the paper's own terms, the discovery is that reconstructed geometry should not be used as direct forensic evidence but as selectively exploited auxiliary information. GFrame reconstructs depth and surface normals from the input, derives confidence-inspired reliability cues from these maps themselves (depth Laplacian smoothness, normal unit-norm deviation, and depth-normal agreement), and uses them to build a spatial gate that controls the residual injection of geometric features into the RGB feature stream at each encoder stage. A cross-scale query module then propagates coarse structural context to fine resolutions, and a boundary-aware loss emphasizes mask transitions. The paper claims

What carries the argument

The central mechanism is a reliability-gated residual fusion: at each encoder stage, a sigmoid gate computed from RGB, depth, normal features and three hand-crafted confidence cues multiplies a weighted sum of depth and normal features before adding it to the RGB feature stream. The confidence cues are the depth Laplacian smoothness (exp(-|∇²D|)), the deviation of the normal magnitude from unit norm, and the agreement between depth-derived and reconstructed normals. These gates exist to admit geometry only where reconstruction is locally trustworthy, so that geometric noise is suppressed before it contaminates appearance-based learning; the cross-scale query module (CSQM) and boundary-aware

Load-bearing premise

The load-bearing premise is that the three hand-crafted reliability proxies—depth Laplacian smoothness, unit-norm deviation, and depth-normal agreement—actually mark where monocular reconstruction is trustworthy on manipulated images; the paper validates them only indirectly through end-task ablations, never against reconstruction error.

What would settle it

Compare the proxy maps (Eqs. 1–3) to per-pixel reconstruction error between estimated depth/normals and ground-truth geometry on manipulated images. If high-proxy regions do not correspond to low-error regions, the gating mechanism's premise fails. Alternatively, replace the gate with an always-open or random gate; if performance does not drop, selective fusion is not the cause of the reported gains.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • With the same 20-epoch training budget, GFrame outperforms all compared baselines on both in-distribution and out-of-distribution averages, and it keeps improving in-distribution with 50 epochs (0.592 to 0.621).
  • Geometric information contributes even when depth or normals are used alone; removing the whole geometric branch drops in-distribution F1 by about 0.04 and out-of-distribution F1 by about 0.04.
  • The framework remains comparatively strong under Gaussian noise, blur, JPEG compression, and social-media recompression, where local appearance statistics are degraded.
  • Replacing the reconstruction backend changes accuracy, indicating that the reliability of the monocular geometry, not just the fusion design, drives the gain.
  • The proposed boundary-aware loss helps the model reach peak validation performance earlier, suggesting that training objectives matter for absorbing geometric cues.

Where Pith is reading between the lines

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

  • Editorial inference: The paper leaves implicit that the same reliability-gated fusion principle could transfer to other forensic tasks, such as deepfake detection or generalized anomaly localization, where monocular reconstruction is also noisy.
  • Editorial inference: A natural testable extension would be replacing the hand-crafted confidence cues with learned uncertainty estimates from the reconstruction model, which might make the gate more accurate and further improve localization.
  • Editorial inference: The strong out-of-distribution results suggest geometry acts as a domain-invariant cue; combining it with other geometric signals, such as lighting or vanishing points, could push the approach further.
  • Editorial inference: The paper's ablations do not test the reliability proxies directly against reconstruction error; if the proxies are uncorrelated with actual error, the reported gains would need re-interpretation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes GFrame, an image manipulation localization (IML) method that augments RGB features with monocular depth and surface normal maps. To handle the noise of reconstructed geometry, GFrame introduces an Uncertainty-aware Geometric Fusion Module (UGFM) that computes hand-crafted reliability cues (depth Laplacian smoothness, unit-norm deviation, and depth-normal agreement) and uses them, along with the geometric features, to produce a spatial gate that selectively injects geometric information into the RGB stream. A Cross-Scale Query Module (CSQM) propagates coarse structural context to fine resolutions, and a boundary-aware loss with margin regularization is used for training. Experiments on several in-distribution (InD) and out-of-distribution (OOD) datasets, all re-evaluated under a unified benchmark framework, report F1@0.5 improvements over six baselines. Ablations show that removing the geometric branch, UGFM, CSQM, or the proposed loss degrades performance.

Significance. If the central claims hold, this work would provide a practical and transferable way to exploit 3D geometry for IML, a direction that is relatively underexplored. The paper is notable for its systematic benchmarking under a common framework, its ablation of the whole geometric branch, and its explicit acknowledgment that reconstruction noise must be handled selectively. The code-release commitment is also a strength. However, the central mechanism — the reliability cues that gate geometric information — is not directly validated, and the reported quantitative advantages are not accompanied by statistical significance tests. The overall approach is plausible, but the evidence as presented does not yet establish that the specific reliability-gating mechanism, rather than the mere addition of geometric features and extra parameters, is responsible for the gains.

major comments (3)
  1. [§3.1, Eqs. (1)–(6) and Table 3] The paper's core claim is that reconstructed geometry should be exploited selectively, and that the confidence-inspired cues C_d, C_n, A_dn identify where geometry is trustworthy. However, the only UGFM ablation ('w/o UGFM') removes the entire fusion module, not merely the hand-crafted cues. Thus the improvement attributed to reliability gating could instead come from simply concatenating raw depth/normal features, from the added parameters, or from the boundary-aware loss. To support the 'selective' claim, the authors should either (a) directly validate the cues against reconstruction error on manipulated images, or (b) add ablations that keep the UGFM architecture but replace the reliability cues with constants, random values, or learned statistics not derived from the geometric maps. Without such experiments, the connection between Eqs. (1)–(3) and the observed performance gains is no
  2. [§4.1, dataset preparation] The statement 'the NIST16 data used in our experiments are manually cleaned to avoid data leakage' raises a serious validity concern. Manual cleaning of a standard benchmark can introduce subjective bias, and if the cleaning is not applied identically to all baselines, the comparisons are unfair. The paper must specify the exact cleaning criteria, the number of images removed or altered, and confirm that every baseline was evaluated on exactly the same cleaned set. It should also clarify what 'data leakage' means here and why the official NIST16 split is insufficient.
  3. [Tables 1, 3, 4, and 5] All F1 numbers are reported from single runs without error bars, multiple seeds, or significance tests. Given the small test sets (e.g., Coverage has 30 test images; NIST16 181), differences of 0.02–0.04 in averaged F1 may be within run-to-run variation. The central quantitative claim — that GFrame outperforms strong baselines — needs to be supported by at least three seeds with mean±std, or by a paired significance test across datasets. This is particularly important for comparisons where the margin is small (e.g., InD average 0.592 vs. Mesorch-50 0.565).
minor comments (6)
  1. [Table 1] The asterisk on 'GFrame 50*' is not explained in the table caption. Also, PIM has no 50-epoch row; this should be noted in the caption or text.
  2. [Eq. (3)] The derivation of the normal field N_d from the depth map is not described. Please define how the depth-derived normals are computed.
  3. [§4.2] The backbone is written 'PvtV2-B2' — presumably 'PvtV2-B2' (or 'PvT-V2'). Please correct. The name 'IMDLBenco' is also written inconsistently ('IMDLBenCo' elsewhere).
  4. [Figure 3] The validation F1 curves lack a y-axis label and do not specify which validation set is used. Please clarify whether these curves are on an InD or OOD split.
  5. [Eq. (11)] The signed prediction confidence S_p is not defined. If it is the pre-sigmoid logit, please state that explicitly.
  6. [General] The reference [9] lists incomplete author names ('J Hsu and SF Chang'). Please use full author names and the correct publication details.

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained and reliability cues are hand-specified rather than fitted.

full rationale

GFrame's derivation does not reduce to its own inputs. The reliability cues C_d, C_n, and A_dn (Eqs. 1-3) are hand-crafted functions of the reconstructed depth and normal maps and are never fit to ground-truth masks; the paper explicitly disclaims them as exact uncertainty estimates. The learned gate G^i (Eq. 5) is trained on labels, but the claim is about the end-to-end localization system, which is evaluated on held-out InD test splits and OOD datasets against baselines retrained under the same budget. Ablations (Table 3) remove components rather than revealing a fitted parameter renamed as prediction. No equation equals another by construction, and there are no self-citations: the reference list contains no work by the present authors. The skeptic's point that the reliability proxies are never validated against reconstruction error is a legitimate experimental-validity concern, but it is not circularity; a hand-specified heuristic that is not derived from the target labels cannot make the derivation circular. Hence score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No new physical or ontological entities are introduced. The reliability cues C_d, C_n, and A_dn are handcrafted computed features derived from depth/normal maps, not independently postulated entities with external falsifiable handles. The free-parameter burden is concentrated in trained scalars and unreported loss hyperparameters, plus the choice of reconstruction backend.

free parameters (6)
  • w_d, w_n (learnable scalar fusion weights) = learned; values not reported
    Eq. (6) uses these scalars to weight depth and normal features in the fused representation; they are trained on the target masks.
  • tau (learnable temperature in cosine attention) = learned; value not reported
    Eq. (7) scales the cosine similarity; learned during training.
  • lambda (boundary weighting) = unspecified
    Eq. (9) sets the strength of boundary emphasis in the spatial weighting map; the paper does not give its value or a sensitivity study.
  • m (margin in margin loss) = unspecified
    Eq. (11) introduces a margin hyperparameter that penalizes low-confidence predictions near boundaries; no value is reported.
  • loss weights for auxiliary supervisions = 'normal weights' only
    Section 3.3 states auxiliary losses are used with normal weights but does not specify the actual coefficients.
  • reconstruction backend selection = MoGe2 selected as default
    MoGe2 is chosen as the stronger default over Metric3D based on validation/ablation performance; this is an engineering choice tuned to the benchmarks.
axioms (6)
  • domain assumption Well-blended manipulations that preserve local appearance still violate scene geometry that is recoverable from monocular depth/normals.
    Section 1 and Figure 1 motivate the entire approach; the paper provides no independent evidence that such violations are measurable in the evaluated datasets.
  • domain assumption Monocular reconstruction on manipulated images retains enough true scene geometry for those inconsistencies to be detectable despite reconstruction noise.
    Section 3.1 assumes the reconstructed maps preserve geometry where reconstruction succeeds; this is central and unverified directly.
  • domain assumption Depth Laplacian smoothness is a proxy for local reconstruction reliability.
    Eq. (1) defines C_d = exp(-|∇²D|), equating irregular second-order depth fluctuations with unreliability; this proxy is asserted, not validated.
  • domain assumption Deviation of reconstructed surface normals from unit length indicates reconstruction instability.
    Eq. (2) defines C_n from ||N_p|| - 1; the paper assumes normals should be unit vectors and that deviations flag noise.
  • domain assumption Agreement between depth-derived normals and estimated normals indicates mutual reliability.
    Eq. (3) defines A_dn via the dot product; compatibility is assumed to be a reliability signal without direct validation.
  • domain assumption Boundary emphasis and margin regularization improve learning of geometry-aware cues.
    Eqs. (9)-(11) are motivated by the observation that geometric discrepancies concentrate at boundaries; the loss design is an unproven modeling choice backed only by the w/o-Loss ablation.

pith-pipeline@v1.3.0-alltime-deepseek · 11692 in / 10595 out tokens · 106818 ms · 2026-08-01T16:14:33.631105+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of When 2D Cues Fail: Improving Image Manipulation Localization with Reliable 3D Geometry." pith.science (2026). https://pith.science/paper/O4NXH5YU

@misc{pith2026260718040,
  author       = {Pith},
  title        = {Pith review of: When 2D Cues Fail: Improving Image Manipulation Localization with Reliable 3D Geometry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O4NXH5YU}},
  note         = {Machine review of arXiv:2607.18040}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing image manipulation localization (IML) methods rely heavily on 2D forensic cues, such as low-level artifacts, noise traces, and semantic inconsistencies in the manipulated image. While effective in many cases, these cues become much less discriminative when manipulated regions are well blended with their surrounding context in appearance. In such cases, a manipulated region may remain locally appearance-consistent, but still violate the geometric structure of the surrounding scene. This limitation motivates us to go beyond purely 2D evidence and introduce geometric reasoning into IML. To this end, we leverage monocular reconstruction to obtain auxiliary geometric cues, including depth and surface normals. However, a key challenge lies in the fact that reconstructed geometry on manipulated images is inherently noisy and cannot be used naively. Rather than treating depth and normals as direct evidence, we estimate their reliability and exploit them selectively for localization. Based on this principle, we design a geometry-aware framework (GFrame) that fuses reliable geometric cues with RGB features and propagates them across scales to improve fine-grained localization. Extensive experiments show that the proposed method achieves excellent performance under limited budget constraints. These results indicate that reliable 3D geometry provides complementary forensic evidence beyond traditional 2D cues for IML. Related code will be released.

Figures

Figures reproduced from arXiv: 2607.18040 by Dan Ma, Gaobo Yang, Guofeng Yu, Zhiqing Guo.

Figure 1
Figure 1. Figure 1: Even when manipulated regions are visually well [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of GFrame. Given an input RGB image, we first reconstruct depth and surface normal maps as auxiliary [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Validation F1 curves of representative settings. The [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on in-distribution (InD) and out-of-distribution (OOD) samples. Each row corresponds to [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 3 linked inside Pith

  1. [1]

    Tiziano Bianchi and Alessandro Piva. 2012. Image forgery localization via block- grained analysis of JPEG artifacts.IEEE Transactions on Information Forensics and Security7, 3 (2012), 1003–1017

    2012

  2. [2]

    Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. 2022. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE Transactions on Pattern Analysis and Machine Intelligence45, 3 (2022), 3539– 3553

    2022

  3. [3]

    Jing Dong, Wei Wang, and Tieniu Tan. 2013. Casia image tampering detection evaluation database. In2013 IEEE China summit and international conference on signal and information processing. IEEE, 422–426

    2013

  4. [4]

    Pasquale Ferrara, Tiziano Bianchi, Alessia De Rosa, and Alessandro Piva. 2012. Image forgery localization via fine-grained analysis of CFA artifacts.IEEE Trans- actions on Information Forensics and Security7, 5 (2012), 1566–1577

    2012

  5. [5]

    Huazhu Fu and Xiaochun Cao. 2012. Forgery authentication in extreme wide- angle lens using distortion cue and fake saliency map.IEEE Transactions on Information Forensics and Security7, 4 (2012), 1301–1314

    2012

  6. [6]

    Thomas Gloe, Karsten Borowka, and Antje Winkler. 2010. Efficient estimation and large-scale evaluation of lateral chromatic aberration for digital image forensics. InMedia Forensics and Security II, Vol. 7541. SPIE, 62–74

    2010

  7. [7]

    Haiying Guan, Mark Kozak, Eric Robertson, Yooyoung Lee, Amy N Yates, Andrew Delgado, Daniel Zhou, Timothee Kheyrkhah, Jeff Smith, and Jonathan Fiscus

  8. [8]

    Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. 2023. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20606–20615

    2023

  9. [9]

    J Hsu and SF Chang. 2006. Columbia uncompressed image splicing detection evaluation dataset.Columbia DVMM Research Lab6 (2006)

    2006

  10. [10]

    Xuefeng Hu, Zhihan Zhang, Zhenye Jiang, Syomantak Chaudhuri, Zhenheng Yang, and Ram Nevatia. 2020. SPAN: Spatial pyramid attention network for image manipulation localization. InEuropean conference on computer vision. Springer, 312–328

    2020

  11. [11]

    Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. 2018. Fighting fake news: Image splice detection via learned self-consistency. InProceedings of the European conference on computer vision (ECCV). 101–117

    2018

  12. [12]

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. 2020. Analyzing and Improving the Image Quality of StyleGAN. In Proc. CVPR

    2020

  13. [13]

    Chenqi Kong, Anwei Luo, Shiqi Wang, Haoliang Li, Anderson Rocha, and Alex C Kot. 2025. Pixel-inconsistency modeling for image manipulation localization. IEEE Transactions on Pattern Analysis and Machine Intelligence(2025)

    2025

  14. [14]

    Paweł Korus and Jiwu Huang. 2016. Evaluation of random field models in multi- modal unsupervised tampering localization. In2016 IEEE international workshop on information forensics and security (WIFS). IEEE, 1–6

    2016

  15. [15]

    Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. 2022. PSCC-Net: Progres- sive spatio-channel correlation network for image manipulation detection and localization.IEEE Transactions on Circuits and Systems for Video Technology32, 11 (2022), 7505–7517

    2022

  16. [16]

    Xiaochen Ma, Bo Du, Zhuohang Jiang, Ahmed Y Al Hammadi, and Jizhe Zhou

  17. [17]

    Xiaochen Ma, Xuekang Zhu, Lei Su, Bo Du, Zhuohang Jiang, Bingkui Tong, Zeyu Lei, Xinyu Yang, Chi-Man Pun, Jiancheng Lv, et al . 2024. Imdl-benco: A comprehensive benchmark and codebase for image manipulation detection & localization.Advances in Neural Information Processing Systems37 (2024), 134591–134613

    2024

  18. [18]

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. 2021. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741(2021)

    Pith/arXiv arXiv 2021

  19. [19]

    Adam Novozamsky, Babak Mahdian, and Stanislav Saic. 2020. IMD2020: A Large- Scale Annotated Dataset Tailored for Detecting Manipulated Images. In2020 IEEE Winter Applications of Computer Vision Workshops (W ACVW). 71–80

    2020

  20. [20]

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. 2021. Vision transformers for dense prediction. InProceedings of the IEEE/CVF international conference on computer vision. 12179–12188

    2021

  21. [21]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

    2022

  22. [22]

    Lei Su, Xiaochen Ma, Xuekang Zhu, Chaoqun Niu, Zeyu Lei, and Ji-Zhe Zhou

  23. [23]

    Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. 2023. Depth-aided Camouflaged Object Detection. InProceedings of the 31st ACM International Conference on Multimedia (ACM MM). 3297–3306. doi:10.1145/3581783.3611874

    arXiv 2023

  24. [24]

    Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. 2025. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. InProceedings of the Computer Vision and Pattern Recognition Conference. 5261–5271

    2025

  25. [25]

    Ruicheng Wang, Sicheng Xu, Yue Dong, Yu Deng, Jianfeng Xiang, Zelong Lv, Guangzhong Sun, Xin Tong, and Jiaolong Yang. 2025. MoGe-2: Accurate Monoc- ular Geometry with Metric Scale and Sharp Details. arXiv:2507.02546 [cs.CV] https://arxiv.org/abs/2507.02546

    Pith/arXiv arXiv 2025

  26. [26]

    Bihan Wen, Ye Zhu, Ramanathan Subramanian, Tian-Tsong Ng, Xuanjing Shen, and Stefan Winkler. 2016. COVERAGE—A novel database for copy-move forgery detection. In2016 IEEE international conference on image processing (ICIP). IEEE, 161–165

    2016

  27. [27]

    Yue Wu, Wael AbdAlmageed, and Premkumar Natarajan. 2019. Mantra-net: Manipulation tracing network for detection and localization of image forgeries with anomalous features. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9543–9552

    2019

  28. [28]

    Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. 2023. Metric3d: Towards zero-shot metric 3d prediction from a single image. InProceedings of the IEEE/CVF international conference on computer vision. 9043–9053

    2023

  29. [29]

    Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang

  30. [30]

    Xuekang Zhu, Xiaochen Ma, Lei Su, Zhuohang Jiang, Bo Du, Xiwen Wang, Zeyu Lei, Wentao Feng, Chi-Man Pun, and Ji-Zhe Zhou. 2025. Mesoscopic insights: orchestrating multi-scale & hybrid architecture for image manipulation localization. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 11022–11030

    2025

  31. [33]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Pattern-Affinitive Propagation across Depth, Surface Normal and Semantic Segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 7144–7153

  32. [2019]

    In2019 IEEE Winter Applications of Computer Vision Workshops (W ACVW)

    MFC datasets: Large-scale benchmark datasets for media forensic chal- lenge evaluation. In2019 IEEE Winter Applications of Computer Vision Workshops (W ACVW). IEEE, 63–72

  33. [2023]

    IML-ViT: Benchmarking Image Manipulation Localization by Vision Trans- former.arXiv preprint arXiv:2307.14863(2023)

    Pith/arXiv arXiv 2023

  34. [2025]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

    Can we get rid of handcrafted feature extractors? sparsevit: Nonsemantics- centered, parameter-efficient image manipulation localization through spare- coding transformer. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 7024–7032