Pith. sign in

REVIEW 4 major objections 5 minor 24 references

SurGSplat: Progressive Geometry-Constrained Gaussian Splatting for Surgical Scene Reconstruction

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SurGSplat claims that endoscopic 3D reconstruction can be done without SfM initialization by progressively growing 3D Gaussians under projection and depth geometric constraints, and that this yields better novel-view synthesis and pose…

desk verdict Useful incremental step for SfM-free endoscopic 3DGS: NVS gains are real, but the pose-estimation claim needs a rigidity/scale fix and a thicker evaluation before it is taken at face value. read the letter →

arxiv 2506.05935 v2 pith:SZLS6R6A submitted 2025-06-06 cs.GR cs.CV

Yuchao Zheng , Jianing Zhang , Guochen Ning , Hongen Liao This is my paper
classification cs.GRcs.CV
keywords 3DGaussianSplattingsurgicalscenereconstructionendoscopystructure-from-motion-freeposeestimationnovelviewsynthesisgeometricconsistencydepthprior
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

SurGSplat is a method for reconstructing a surgical scene from an unposed endoscopic video. Its central claim is that 3D Gaussian Splatting can be initialized from a single depth map and then grown frame by frame, so the usual Structure-from-Motion preprocessing step is unnecessary and its failure modes, such as sparse texture and inconsistent lighting, are avoided. Two geometric losses, one enforcing consistency between matched pixels and one correlating rendered depth with a pretrained depth prior, keep the growing Gaussian scene aligned with the video. On four scenes from the SCARED porcine dataset, the authors report that SurGSplat improves novel-view synthesis and pose estimation over NoPe-NeRF and CF-3DGS, while COLMAP fails on several scenes. A sympathetic reader would take the contribution to be an SfM-free 3DGS pipeline tailored to endoscopic conditions.

What carries the argument

The central mechanism is the progressive local-to-global growth of the 3D Gaussian scene: first-frame depth initializes the Gaussians, each new frame's pose is estimated by optimizing an affine transformation of Gaussian centers, and the accumulated poses drive densification and global refinement. The two named constraints are the Projection Geometric Consistency Loss, which compares dense feature-matching distances between consecutive frames with the re-projection distances induced by the estimated poses, and the Depth Geometric Correlation Loss, which computes local Pearson correlation between rendered depth patches and monocular depth prior patches. Together they supply 2D-3D geometric supervision that keeps the growing reconstruction aligned without any SfM preprocessing.

What would settle it

Run SurGSplat on a long endoscopic sequence with ground-truth camera poses and monitor absolute trajectory error as frames are added: if the error grows without bound or spikes sharply after a single corrupted depth frame, then the assumption that monocular depth provides a stable geometric scaffold is broken.

Watch

Extended reading notes

Core claim

The paper argues that 3D Gaussian Splatting can serve as a complete SfM-free reconstruction engine for endoscopic video if the scene is grown progressively and supervised by geometric consistency. SurGSplat seeds the global Gaussian set from the first frame's monocular depth, then per frame estimates the relative camera motion as an affine transformation of Gaussian centers using only photometric loss, accumulates those relative poses into a trajectory, and finally updates the global Gaussians with visual, projection geometric consistency, and depth geometric correlation losses. The authors state that this design removes the dependency on SfM point clouds and poses, which routinely fail in low-texture surgical scenes, and that the geometric constraints preserve fine structures such as blood vessels. Their reported quantitative results on SCARED show higher PSNR and SSIM, lower LPIPS, and lower translation and rotation errors than the SfM-free baselines, including scenes where COLMAP produced no reconstruction at all.

Load-bearing premise

The method's accuracy rests on the pretrained monocular depth maps being a trustworthy geometric scaffold across all frames, with no correction for overall scale and no global step that ties all frames together.

Editorial extensions

If this is right

  • Endoscopic video reconstruction no longer requires a successful COLMAP run, removing a common failure point in low-texture surgical scenes.
  • The method produces dense 3D Gaussians with fine vascular detail, providing a practical basis for intraoperative navigation and lesion localization.
  • Because pose estimation and Gaussian growing share one optimization loop on a single GPU, the pipeline is fast enough to fit into intraoperative workflows.
  • The progressive growth design extends naturally to longer videos, allowing scene coverage to expand over time rather than being reconstructed offline from a fixed frame set.
  • The stated scope is dense endoscopic video; sparse-view preoperative reconstruction is left as future work by the authors.

Reading between the lines

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

  • One testable extension is to add a global bundle-adjustment or metric-scale alignment step and measure whether absolute trajectory error stays bounded on sequences far longer than the 40-frame clips used here.
  • Because the depth correlation loss is insensitive to absolute depth scale, the recovered geometry may be correct only up to scale; coupling the method with known endoscope kinematics or stereo frames could pin down absolute scale.
  • The same progressive geometric-consistency recipe could be adapted to non-rigid tissue by replacing the affine warp with a per-local deformation field, addressing breathing motion or tool-tissue interaction that the current rigid-pose assumption does not cover.
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. The paper proposes SurGSplat, a Structure-from-Motion-free 3D Gaussian Splatting pipeline for endoscopic scene reconstruction. The method initializes a global Gaussian representation from a single monocular depth map, estimates relative camera poses by optimizing an affine transformation of local Gaussians against the photometric loss, progressively grows the global Gaussian set, and refines it with two geometric losses: a Projection Geometric Consistency loss based on dense 2D matching distances and a Depth Geometric Correlation loss based on patch-wise Pearson correlation of rendered and monocular depth maps. Experiments on four scenes of the SCARED dataset compare SurGSplat with COLMAP+3DGS, NoPe-NeRF, and CF-3DGS, reporting improved novel-view synthesis metrics and pose error values comparable to or slightly better than CF-3DGS.

Significance. If the claims are validated, the paper addresses a real clinical bottleneck: SfM failure in low-texture, poorly lit endoscopic scenes. The use of public data (SCARED) and external pretrained models (Depth Anything, RoMa) is a strength, as is the explicit goal of eliminating SfM preprocessing. However, the current evidence is insufficient to support the central claims of accurate pose estimation and metric reconstruction. The pose optimization is unconstrained in a way that may not correspond to camera motion, the depth priors are scale-ambiguous and not aligned to metric ground truth, and the reported pose improvements over CF-3DGS are marginal. The lack of ablations, error bars, and key hyperparameter values further limits the reproducibility and attribution of gains to the proposed geometric losses. The idea is plausible and potentially useful, but the experimental and theoretical support needs substantial strengthening.

major comments (4)
  1. [II-B2, Eqs. (5)-(7)] The claim that estimating camera pose T_t is 'mathematically analogous' to optimizing an affine transformation A_t is not established. For a pinhole camera, a 3D affine transformation of the scene followed by a fixed camera projection is equivalent to a camera motion only when the affine map is a similarity (a scaled rigid transformation). The optimization in Eq. (7) does not constrain A_t to SE(3) or even to a similarity, so the photometric loss may converge to non-rigid affine solutions that render the next frame well but do not correspond to any valid camera pose. The paper needs to either restrict A_t to rigid/similarity transformations, add a rigidity regularizer, or demonstrate empirically that the optimized A_t is close to a rigid transformation.
  2. [II-C, Eqs. (10)-(11)] The scale ambiguity of the monocular depth prior is not resolved, so the reported pose errors in Table I are not directly comparable to metric ground truth. Depth Anything provides depth up to per-image scale and shift; the DGC loss in Eq. (11) is explicitly a normalized Pearson correlation that is invariant to affine depth transforms, and the PGC loss in Eq. (10) compares only 2D pixel distances. No metric scale alignment, Sim(3) alignment, or global bundle adjustment is described anywhere in the method or evaluation. Without disclosing an alignment procedure before computing ATE/RPE, the small absolute values such as ATE=0.008 cannot be interpreted as metric accuracy.
  3. [Table I] The claimed pose-estimation superiority over CF-3DGS is not visible in the numbers. In Scene 2, SurGSplat has RPE_trans 1.013 versus 1.012 for CF-3DGS; in Scene 5, 1.531 versus 1.525; ATE differences are at the fourth decimal place. Only in Scene 1 is there a marginal improvement (0.476 vs. 0.478). No error bars, standard deviations, or repeated-run statistics are provided. The abstract's statement of 'superior performance in both novel view synthesis and pose estimation accuracy' is therefore not supported by the pose results as presented.
  4. [III-A2 and III-B] No ablation study isolates the contribution of the proposed components. The paper reports results for the full SurGSplat method only, so the effect of the PGC loss, the DGC loss, and the depth correction mask M_t cannot be assessed. In addition, key hyperparameters are not specified: the weights lambda_rgb, lambda_pgc, lambda_dgc, the depth correction threshold k, the DGC patch size a and number of patches N, the confidence threshold for PGC, and the frame/densification interval are all absent from the implementation details. This prevents replication and makes it impossible to determine whether the geometric constraints, rather than other training choices, drive the reported improvements.
minor comments (5)
  1. [III-A1] The text says '40 consecutive frames were randomly selected from each scene'; random selection and 'consecutive' are contradictory. Please clarify how the frames were actually chosen and how the 30/10 train/test split was performed.
  2. [II-B1, Eq. (3)] In Eq. (3) the optimization variables are listed as c0, r0, s0, alpha0, but not mu0. It should be stated explicitly whether the Gaussian centers are fixed to the back-projected depth points or also optimized during initialization.
  3. [II-C, Eq. (11)] The DGC formula as written is not the standard Pearson correlation: the denominator uses (sigma(d)+epsilon)*(sigma(d_hat)+epsilon) instead of sigma(d)*sigma(d_hat). If epsilon is nonzero, this changes the loss. Please define the exact form or use the standard formulation with epsilon added inside the covariance or variance terms.
  4. [Fig. 3 and Table I] COLMAP+3DGS is reported only for Scene 1, and the table header says 'four representative endoscopic scenes' without identifying which SCARED scenes were used. Please state the scene identifiers and explain why COLMAP results are absent for Scenes 2-5.
  5. [IV] The discussion of clinical applicability is qualitative and does not connect to a downstream task metric; sentences such as 'supports surgeons in navigating complex anatomical structures' would be better supported by a quantitative evaluation on a navigation or segmentation proxy.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SurGSplat's geometric losses are supervised by external pretrained priors, not by its own outputs, and its self-citations are background only.

full rationale

The derivation chain is self-contained with respect to the paper's central claims. The initial 3D Gaussian set is built from monocular depth produced by the external Depth Anything model, relative poses are optimized by photometric loss in Eq. (7), and the PGC and DGC losses in Eqs. (10)-(11) use RoMa 2D matches and Depth Anything depth patches as supervision. These are external inputs or priors, not quantities derived from the target reconstruction or pose results, so the reported NVS and pose numbers are not forced by construction. Evaluation is carried out against SCARED ground-truth trajectories and held-out views, providing an external benchmark. The only self-citations, references [5] and [9], are background citations for augmented reality and 3DGS efficiency and are not load-bearing for the method's novelty or correctness. The unconstrained affine transformation in Eqs. (5)-(7) and the scale invariance of the DGC loss raise identifiability and evaluation-alignment concerns, but these are correctness risks rather than circular derivations. The manuscript's own limitation paragraph about sparse-view applicability is a scoping statement, not a circular step.

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

The central claim relies on externally supplied priors (monocular depth and learned dense matching) and on several unstated hyperparameters. No new physical entities are introduced. The main burden is the assumed validity of depth and correspondence priors in low-texture endoscopy, plus the unidentified loss weights and thresholds that are needed to reproduce the results.

free parameters (5)
  • loss weights lambda_rgb, lambda_pgc, lambda_dgc = not reported
    These weights balance the photometric, projection consistency, and depth correlation losses in Eq. (12); no values or tuning procedure are given.
  • depth correction threshold k = not reported
    The depth correction mask M_t clamps rendered depths above a predefined threshold k, but the threshold value is not specified.
  • DGC patch size a and number of patches N = not reported
    Eq. (11) samples N random patches of size a by a from depth maps; neither a nor N is given.
  • PGC confidence threshold = 20%
    Matched points with confidence below 20% are excluded from the projection loss; this threshold is chosen without justification and acts as a data exclusion rule.
  • frame and densification interval = not reported
    The interval for adding new frames is synchronized with point densification, but the actual schedule is not described.
assumptions (4)
  • domain assumption Monocular depth maps from Depth Anything provide a sufficiently consistent geometric scaffold for back-projecting pixels into 3D.
    Used to initialize 3D Gaussians from the first frame and to supervise the DGC loss in Sections II-B and II-C; no metric-scale alignment across frames is described.
  • domain assumption RoMa dense matches delta_s are reliable ground-truth correspondence distances for consecutive endoscopic frames.
    The PGC loss in Eq. (10) drives projected geometry toward these matching distances; if matches fail in low-texture tissue, the loss anchors the reconstruction to incorrect correspondences.
  • domain assumption Estimating an affine transformation A_t of Gaussian centers is equivalent to estimating the camera pose T_t under projection.
    Pose estimation is cast as optimizing A_t in Eq. (7) using the equivalence in Eqs. (5) and (6); the equivalence holds only under rigid or scaled congruence and no pose prior.
  • standard math The standard 3D Gaussian Splatting rendering and alpha-blending equations hold in endoscopic conditions.
    Section II-A introduces Eqs. (1) and (2) as inherited from the original 3DGS framework; this modeling assumption transfers to surgical scenes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SurGSplat: Progressive Geometry-Constrained Gaussian Splatting for Surgical Scene Reconstruction." pith.science (2026). https://pith.science/paper/SZLS6R6A

@misc{pith2026250605935,
  author       = {Pith},
  title        = {Pith review of: SurGSplat: Progressive Geometry-Constrained Gaussian Splatting for Surgical Scene Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SZLS6R6A}},
  note         = {Machine review of arXiv:2506.05935}
}
read the original abstract

Intraoperative navigation relies heavily on precise 3D reconstruction to ensure accuracy and safety during surgical procedures. However, endoscopic scenarios present unique challenges, including sparse features and inconsistent lighting, which render many existing Structure-from-Motion (SfM)-based methods inadequate and prone to reconstruction failure. To mitigate these constraints, we propose SurGSplat, a novel paradigm designed to progressively refine 3D Gaussian Splatting (3DGS) through the integration of geometric constraints. By enabling the detailed reconstruction of vascular structures and other critical features, SurGSplat provides surgeons with enhanced visual clarity, facilitating precise intraoperative decision-making. Experimental evaluations demonstrate that SurGSplat achieves superior performance in both novel view synthesis (NVS) and pose estimation accuracy, establishing it as a high-fidelity and efficient solution for surgical scene reconstruction. More information and results can be found on the page https://surgsplat.github.io/.

Figures

Figures reproduced from arXiv: 2506.05935 by the authors.

Figure 1
Figure 1. Pipeline of SurGSplat for high-fidelity surgical scene reconstruction. Starting with an initial depth map generated from a single image I0, our method iteratively optimizes camera poses and global 3D Gaussian representations. Each frame is processed by constructing a local 3D Gaussian set and optimizing the relative pose via neural rendering. The accumulated poses are used to refine the global 3D Gaussians based on … view at source ↗
Figure 2
Figure 2. Principle of Projection Geometric Consistency Loss (PGC). The matching distance ∆s = M(It−1, It) serves as ground truth to optimize the projective distance ∆ˆs between corresponding 3D points projected from consecutive frames, ensuring geometric consistency across matched features. The unique characteristics of endoscopic imagery, such as sparse textures and intricate structures, pose significant challenges for reli… view at source ↗
Figure 3
Figure 3. Rendering results for four representative scenes in the SCARED dataset. Notably, COLMAP only successfully reconstructed Scene 1, while failing [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 17 canonical work pages

  1. [1]

    Free-DyGS: Camera-Pose-Free Scene Reconstruction for Dynamic Surgical Videos with Gaussian Splatting

    Q. Li, S. Yang, D. Shen, and Y . Jin, “Free-dygs: Camera-pose-free scene reconstruction based on gaussian splatting for dynamic surgical videos,” 2024. [Online]. Available: https://arxiv.org/abs/2409.01003

    work page Pith review arXiv 2024

  2. [2]

    Nr-slam: Nonrigid monocular slam,

    J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard´os, “Nr-slam: Nonrigid monocular slam,”IEEE Transactions on Robotics, vol. 40, pp. 4252– 4264, 2024

    work page 2024

  3. [3]

    Augmented reality navigation for liver resection with a stereoscopic laparoscope,

    H. Luo, D. Yin, S. Zhang, D. Xiao, B. He, F. Meng, Y . Zhang, W. Cai, S. He, W. Zhang, Q. Hu, H. Guo, S. Liang, S. Zhou, S. Liu, L. Sun, X. Guo, C. Fang, L. Liu, and F. Jia, “Augmented reality navigation for liver resection with a stereoscopic laparoscope,” Computer Methods and Programs in Biomedicine, vol. 187, p. 105099, 2020. [Online]. Available: https...

    work page 2020

  4. [4]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” inEuropean Conference on Computer Vision (ECCV), 2020

    work page 2020

  5. [5]

    Augmented reality technology for preoperative planning and intraoperative navigation during hepatobiliary surgery: A review of current methods,

    R. Tang, L.-F. Ma, Z.-X. Rong, M.-D. Li, J.-P. Zeng, X.-D. Wang, H.-E. Liao, and J.-H. Dong, “Augmented reality technology for preoperative planning and intraoperative navigation during hepatobiliary surgery: A review of current methods,”Hepatobiliary & Pancreatic Diseases International, vol. 17, no. 2, pp. 101–112, 2018. [Online]. Available: https://www....

    work page 2018

  6. [6]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, July 2023. [Online]. Available: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

    2023

  7. [7]

    2d gaussian splatting for geometrically accurate radiance fields,

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2d gaussian splatting for geometrically accurate radiance fields,” inSIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024

    2024

  8. [8]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,

    D. Chen, H. Li, W. Ye, Y . Wang, W. Xie, S. Zhai, N. Wang, H. Liu, H. Bao, and G. Zhang, “Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,”arXiv preprint arXiv:2406.06521, 2024

    arXiv 2024

Show all 24 references
  1. [9]

    Gbr: Generative bundle refinement for high-fidelity gaussian splatting and meshing,

    J. Zhang, Y . Zheng, Z. Li, Q. Dai, and X. Yuan, “Gbr: Generative bundle refinement for high-fidelity gaussian splatting and meshing,”

  2. [10]

    Pixel- wise view selection for unstructured multi-view stereo,

    J. L. Sch ¨onberger, E. Zheng, M. Pollefeys, and J.-M. Frahm, “Pixel- wise view selection for unstructured multi-view stereo,” inEuropean Conference on Computer Vision (ECCV), 2016

    2016

  3. [11]

    Structure-from-motion revisited,

    J. L. Sch ¨onberger and J.-M. Frahm, “Structure-from-motion revisited,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  4. [12]

    3d reconstruction from endoscopy images: A survey,

    Z. Yang, J. Dai, and J. Pan, “3d reconstruction from endoscopy images: A survey,”Computers in Biology and Medicine, vol. 175, p. 108546, 2024. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0010482524006309

    2024

  5. [13]

    Tracking and mapping in medical computer vision: A review,

    A. Schmidt, O. Mohareri, S. DiMaio, M. C. Yip, and S. E. Salcudean, “Tracking and mapping in medical computer vision: A review,”Medical Image Analysis, vol. 94, p. 103131, 2024. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S1361841524000562

    2024

  6. [14]

    iNeRF: Inverting neural radiance fields for pose estimation,

    L. Yen-Chen, P. Florence, J. T. Barron, A. Rodriguez, P. Isola, and T.- Y . Lin, “iNeRF: Inverting neural radiance fields for pose estimation,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021

    2021

  7. [15]

    Nerf– : Neural radiance fields without known camera parameters,

    Z. Wang, S. Wu, W. Xie, M. Chen, and V . A. Prisacariu, “Nerf– : Neural radiance fields without known camera parameters,”arXiv preprint arXiv:2102.07064, 2021

    2021 arXiv

  8. [16]

    Barf: Bundle- adjusting neural radiance fields,

    C.-H. Lin, W.-C. Ma, A. Torralba, and S. Lucey, “Barf: Bundle- adjusting neural radiance fields,” inIEEE International Conference on Computer Vision (ICCV), 2021

    2021

  9. [17]

    Nope-nerf: Optimising neural radiance field with no pose prior,

    W. Bian, Z. Wang, K. Li, J. Bian, and V . A. Prisacariu, “Nope-nerf: Optimising neural radiance field with no pose prior,” 2023

    2023

  10. [18]

    Colmap-free 3d gaussian splatting,

    Y . Fu, S. Liu, A. Kulkarni, J. Kautz, A. A. Efros, and X. Wang, “Colmap-free 3d gaussian splatting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 20 796–20 805

    2024

  11. [19]

    Stereo correspondence and reconstruction of endoscopic data challenge,

    M. Allan, J. Mcleod, C. Wang, J. C. Rosenthal, Z. Hu, N. Gard, P. Eisert, K. X. Fu, T. Zeffiro, W. Xia, Z. Zhu, H. Luo, F. Jia, X. Zhang, X. Li, L. Sharan, T. Kurmann, S. Schmid, R. Sznitman, D. Psychogyios, M. Azizian, D. Stoyanov, L. Maier-Hein, and S. Speidel, “Stereo corre...

    2021 arXiv

  12. [20]

    Depth anything: Unleashing the power of large-scale unlabeled data,

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  13. [21]

    RoMa: Robust Dense Feature Matching,

    J. Edstedt, Q. Sun, G. B ¨okman, M. Wadenb ¨ack, and M. Fels- berg, “RoMa: Robust Dense Feature Matching,”Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  14. [22]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Trans- actions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  15. [23]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  16. [2024]

    Available: https://arxiv.org/abs/2412.05908

    [Online]. Available: https://arxiv.org/abs/2412.05908

    arXiv

Pith tools

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