Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Novel View Synthesis with Gaussian Splatting: Impact on Photogrammetry Model Accuracy and Resolution

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper reports that photogrammetry models built from photo sets augmented with Gaussian-Splatting-rendered views score higher on SSIM and PSNR than models from the original photos alone, at the cost of lower measured resolution.

desk verdict A small, honest applied study whose main augmentation claim is undercut by a circular evaluation: the GS-rendered views come from a model trained on the very images used as ground truth. read the letter →

arxiv 2508.07483 v1 pith:CQWTNLIK submitted 2025-08-10 cs.CV eess.IV

Pranav Chougule This is my paper
classification cs.CVeess.IV
keywords Photogrammetry3DGaussianSplattingNovelviewsynthesisDatasetaugmentationSSIMPSNRLPIPSUSAFresolutionchart
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tests whether synthetic images produced by 3D Gaussian Splatting — a scene represented as learnable 3D Gaussian blobs rasterized into images — can feed back into photogrammetry and improve its output. The author captured indoor and outdoor photo sets, built 3D models with both photogrammetry and Gaussian Splatting, and modified the Gaussian Splatting codebase so it renders images from arbitrary camera poses defined in Blender instead of only the original capture poses. Adding those rendered novel views to the original photos produced an augmented photogrammetry model that scored higher on SSIM and PSNR than the photos-only model in both scenes. The synthetic views also improved scene completeness, but views rendered from poses far from the original cameras introduced noise that lowered the model's measured lp/mm resolution. The paper positions synthetic view synthesis as a practical way to densify input coverage for photogrammetric reconstruction, with novel pose placement as the key quality control.

What carries the argument

The load-bearing element is a modified 3D Gaussian Splatting renderer. Gaussian Splatting models a scene as a collection of learnable 3D Gaussian functions — position, shape, opacity, color — that are rasterized into images; the paper forks an existing implementation of this method and removes its dependence on the original Structure-from-Motion camera poses (the images.bin and cameras.bin files), replacing them with user-defined camera poses generated in Blender on circular paths around the photogrammetry mesh and converted to the renderer's pose format with the camorph library. This is what makes novel view synthesis from arbitrary viewpoints possible. The second mechanism is the augmentat

What would settle it

Split one dataset's ground-truth images in two; train Gaussian Splatting on half A, render novel views of half B's viewpoints, add the renders to half A for photogrammetry, and evaluate the augmented model on half B alone. If SSIM/PSNR gains persist, augmentation adds real information; if they vanish, the reported gains come from test-set leakage. Second check: render novel views at increasing radial offsets from the original cameras and plot measured lp/mm against offset to find the pose distance at which augmentation starts reducing resolution.

Watch

Extended reading notes

Core claim

The central claim: a photogrammetry model reconstructed from a dataset augmented with Gaussian-Splatting-rendered novel views is more accurate, by SSIM and PSNR, than a model built from the original ground-truth images alone, in both indoor and outdoor scenes. Gaussian Splatting by itself also beats photogrammetry on SSIM, PSNR, and LPIPS at the original poses. The enabling result is a modification of an existing Gaussian Splatting codebase that decouples rendering from the Structure-from-Motion poses in images.bin and cameras.bin, so any Blender-defined camera path, converted with camorph, can be rendered. It also documents trade-offs: the augmented model gains completeness but loses resolu

Load-bearing premise

The Gaussian Splatting novel views added to the photogrammetry dataset (Section II-D) are rendered from a model trained on the same ground-truth images that later serve as the test set for the augmented model (Section III-B), so the reported SSIM/PSNR gains could come from the model having seen those viewpoints indirectly rather than from genuinely new scene information.

Editorial extensions

If this is right

  • Photogrammetry models built from datasets augmented with Gaussian-Splatting-rendered novel views scored higher on SSIM and PSNR than models built from the original photos alone, in both the indoor and outdoor scenes tested (Section III-B).
  • Gaussian Splatting by itself outperformed photogrammetry on SSIM, PSNR, and LPIPS at the original camera poses for both datasets.
  • The augmentation gain carries a resolution cost: the indoor augmented model measured 0.445 lp/mm versus 0.561 lp/mm for the original model on the USAF 1951 chart, because splatted views from poses far from the original cameras are noisy.
  • Novel camera pose placement is therefore the controlling design factor; the paper's own next step is to replace circular camera paths with poses at controlled offsets on all sides of the original cameras.
  • For applications such as extended reality and autonomous-vehicle simulation, the demonstrated pipeline shows a route to densify camera coverage synthetically before a photogrammetry pass.

Reading between the lines

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

  • Editorial inference: the evaluation is not fully independent. The Gaussian-Splatting novel views are rendered from a model trained on the same ground-truth images that later serve as the test set for the augmented photogrammetry model, so the added views carry no scene information beyond the test set; part of the SSIM/PSNR gain may come from the augmented model having seen the test viewpoints indi
  • Editorial inference: the claim could be settled by holding a subset of ground-truth images out of both the Gaussian Splatting training and the photogrammetry reconstruction, rendering novel views only from the remaining images, and scoring the augmented model on the held-out set — a protocol the paper does not run.
  • Editorial inference: if the gain is mostly viewpoint leakage, inserting NeRF- or diffusion-rendered views through the same protocol should reproduce the SSIM/PSNR pattern; if the gain is genuine added coverage, the effect size should track the renderer's fidelity, giving a direct test of whether the effect is specific to Gaussian Splatting.
  • Editorial note: the photogrammetry mesh-and-texture workflow described in Section II-B is cited to a document that is itself a citation-format guide, so that procedural description lacks a supporting technical reference; replication should verify the reconstruction settings directly.
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 presents an applied comparison of photogrammetry and 3D Gaussian Splatting (3DGS) on two real-world datasets (indoor and outdoor). The author modifies an existing 3DGS repository to render images from user-defined camera poses generated in Blender, and uses those rendered novel views to create an augmented dataset for photogrammetry. The central claims are (i) 3DGS outperforms photogrammetry in SSIM/PSNR/LPIPS when evaluated on the original camera poses, and (ii) photogrammetry models trained on the augmented dataset (original images plus 3DGS novel views) outperform models trained on original images alone in SSIM and PSNR. Resolution is assessed with a USAF 1951 chart. The paper concludes that 3DGS-generated novel views can improve photogrammetry-based reconstruction.

Significance. If supported, the paper would offer practical guidance for using 3DGS novel views to augment photogrammetry datasets, and it contributes an engineering artifact (modified 3DGS rendering pipeline) with released code. The use of real indoor and outdoor scenes, multiple image-quality metrics, and a resolution chart is commendable. However, the experimental design has a load-bearing circularity: the novel views added to the augmented dataset are generated by a 3DGS model trained on the same original images used to evaluate the augmented photogrammetry model. The reported SSIM/PSNR gains therefore do not establish that augmentation improves generalization to unseen viewpoints. In addition, the 3DGS-vs-photogrammetry comparison is evaluated on training views, and the augmentation comparison confounds dataset size with augmentation content. These issues must be resolved before the central claims can be accepted.

major comments (4)
  1. [Section II-D / III-B] The augmented dataset is built from novel views rendered by a 3DGS model trained on the original ground-truth images, and those same original images are then used as the evaluation set for the augmented photogrammetry model. The synthetic views are deterministic functions of the test images, so any SSIM/PSNR improvement may reflect information leakage rather than better reconstruction of unseen viewpoints. This directly undermines the central claim that GS-augmented photogrammetry 'consistently outperform[s]' the original. To support the claim, hold out a subset of original images for evaluation only, train both the 3DGS renderer and the photogrammetry models exclusively on the remaining images, and then compare on the held-out views.
  2. [Section III-A] The comparison between 3DGS and photogrammetry is performed on the original camera poses used to train the 3DGS model. 3DGS optimizes its Gaussians to reproduce exactly those training views, so higher SSIM/PSNR is expected and does not demonstrate superior novel-view synthesis. Please either evaluate on held-out real views or explicitly reframe Section III-A as a comparison of training-view reconstruction fidelity rather than novel-view generalization.
  3. [Section III-B] The augmented photogrammetry model is trained with 325 images while the original model uses only 225 images, so any improvement could be an effect of dataset size rather than the specific GS-rendered content. A control augmentation is needed, for example adding the same number of images generated by a non-GS baseline (e.g., simple image warps or additional real images from nearby poses) to isolate the contribution of the GS novel views. Without this control, the attribution to Gaussian Splatting is not established.
  4. [Section III-B / Conclusion] The claim that augmented models 'consistently' outperform original models is based on one indoor scene and one outdoor scene, each with a single run and no error bars or significance testing. This is insufficient for a general conclusion. At minimum, report multiple trials with variance; ideally, add more scenes to support the claimed consistency.
minor comments (5)
  1. [General] Tables 1–5 are referenced in the text but are not present in the submitted manuscript. The numerical results for the central comparisons cannot be verified; please include all tables.
  2. [Section II-E, Eq. (4)] The USAF resolution formula is garbled and incomplete. Please provide the standard equation relating group/element number to line pairs per millimeter, and explain how lp/mm was measured from the rendered images.
  3. [References] Reference [5] is listed as Mildenhall et al., NeRF, but it is cited for 'Multiview Neural Surface Reconstruction by Srinivasan et al.'; the citation appears mismatched. Reference [4] is cited as DeepView but lists 'Deepstereo'; please correct these.
  4. [General] There are numerous typographical/encoding artifacts (e.g., 'poison mesh' for Poisson mesh, corrupted mathematical variables). A careful proofread is needed.
  5. [Section III-D] The augmented photogrammetry model has lower USAF resolution (0.445 lp/mm vs 0.561 lp/mm). This is acknowledged, but the title promises 'Accuracy and Resolution'; the paper should clarify the trade-off between improved SSIM/PSNR and degraded resolution.

Circularity Check

1 steps flagged · score 6.0 of 10

GS-augmented photogrammetry is evaluated in a closed loop: the added novel views are rendered from a model trained on the same original images used as ground truth, so the reported SSIM/PSNR gains are not an independent validation.

  1. fitted input called prediction [II-B (GS training), II-D (augmented dataset creation), III-B (comparison)]
    "The Gaussian splats undergo training with a neural network using Stochastic Gradient Descent to optimize their parameters, minimizing the loss between the rasterized and actual images. ... An augmented dataset was created by combining the original ground images with novel views synthesized using the Gaussian Splatting method."

    The GS model that renders the novel views is fitted to the same original ground-truth images that later serve as the evaluation reference for the photogrammetry models. The augmented dataset (II-D) is simply the union of those originals with GS-rendered views, so the additional training images are deterministic functions of the test/evaluation set. When Section III-B reports that the augmented photogrammetry model achieves higher SSIM/PSNR than the original model, it is comparing models on the very images from which the added synthetic views were generated. The augmented model has seen extra views that contain no scene information beyond the evaluation images; any improvement can therefore be explained by training on test-derived data rather than by genuine novel-view benefit. The 'predict

full rationale

The central claim—that photogrammetry models built from GS-augmented datasets outperform those built from original images in SSIM/PSNR—relies on an augmented dataset whose added images are produced by a Gaussian Splatting model trained on the same original ground-truth images used as the comparison reference. Because no held-out views are used, the evaluation cannot distinguish genuine reconstruction improvement from the trivial effect of adding training examples that are functions of the test set. This is a data-independence leak rather than a logical self-definition: the paper does not define GS quality in terms of photogrammetry accuracy, and there are no load-bearing self-citations or imported uniqueness theorems. Still, the primary empirical conclusion reduces to a comparison on training-derived data, so a moderate circularity score is warranted. The resolution comparison (Section III-D) actually reports a resolution decrease and is not part of the circular loop.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its empirical claims rest on standard pipelines and on the assumption that synthetic GS views are independent scene measurements, which is not the case.

free parameters (3)
  • Number of GS-rendered novel views = 100 (indoor dataset)
    Chosen by the author; the augmented dataset used 100 GS images with 225 originals. The count affects the balance and likely the results.
  • Blender camera pose ring radius and elevation = Not reported numerically
    Camera poses were placed on circular paths with customizable radius and elevation; the paper notes that poses far from original views introduce noise, so these choices directly affect quality.
  • GS training hyperparameters = Not reported
    Training on Google Colab T4 with unspecified iterations and learning rates; these affect rendered view quality.
assumptions (4)
  • domain assumption COLMAP and the Gaussian Splatting training pipeline produce accurate camera poses and reconstructions for both datasets.
    The entire comparison relies on the correctness of these standard tools.
  • domain assumption GS-rendered novel views are treated as valid scene observations for photogrammetry even though they are generated from a model trained on the original ground truth images.
    This is the load-bearing premise for the augmentation comparison; it is questionable because the views contain no new information.
  • domain assumption SSIM, PSNR, LPIPS computed on original camera poses are appropriate indicators of 3D reconstruction quality.
    The paper evaluates rendering quality at training viewpoints, which favors methods that overfit to those views.
  • domain assumption The USAF 1951 chart resolution measurement reflects the true spatial resolution of the reconstructed models.
    Manual identification of resolved elements is subjective and depends on rendering viewpoint.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Novel View Synthesis with Gaussian Splatting: Impact on Photogrammetry Model Accuracy and Resolution." pith.science (2026). https://pith.science/paper/CQWTNLIK

@misc{pith2026250807483,
  author       = {Pith},
  title        = {Pith review of: Novel View Synthesis with Gaussian Splatting: Impact on Photogrammetry Model Accuracy and Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CQWTNLIK}},
  note         = {Machine review of arXiv:2508.07483}
}
read the original abstract

In this paper, I present a comprehensive study comparing Photogrammetry and Gaussian Splatting techniques for 3D model reconstruction and view synthesis. I created a dataset of images from a real-world scene and constructed 3D models using both methods. To evaluate the performance, I compared the models using structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), learned perceptual image patch similarity (LPIPS), and lp/mm resolution based on the USAF resolution chart. A significant contribution of this work is the development of a modified Gaussian Splatting repository, which I forked and enhanced to enable rendering images from novel camera poses generated in the Blender environment. This innovation allows for the synthesis of high-quality novel views, showcasing the flexibility and potential of Gaussian Splatting. My investigation extends to an augmented dataset that includes both original ground images and novel views synthesized via Gaussian Splatting. This augmented dataset was employed to generate a new photogrammetry model, which was then compared against the original photogrammetry model created using only the original images. The results demonstrate the efficacy of using Gaussian Splatting to generate novel high-quality views and its potential to improve photogrammetry-based 3D reconstructions. The comparative analysis highlights the strengths and limitations of both approaches, providing valuable information for applications in extended reality (XR), photogrammetry, and autonomous vehicle simulations. Code is available at https://github.com/pranavc2255/gaussian-splatting-novel-view-render.git.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. SwiftGS: Episodic Priors for Immediate Satellite Surface Recovery

    cs.CV 2026-03 unverdicted novelty 6.0 of 10

    SwiftGS uses episodic meta-training to predict geometry-radiation-decoupled Gaussian primitives and a lightweight SDF for zero-shot 3D satellite surface reconstruction with physics-aware rendering.

  2. SwiftGS: Episodic Priors for Immediate Satellite Surface Recovery

    cs.CV 2026-03 conditional novelty 6.0 of 10

    SwiftGS predicts satellite 3D surfaces and renderings zero-shot via meta-learned Gaussian-SDF hybrid, reporting 1.22 m DSM MAE on DFC2019 at 2.5 min per scene.

Reference graph

Works this paper leans on

16 extracted references · 14 canonical work pages · cited by 1 Pith paper

  1. [1]

    Teza, Giordano, Arianna Pesci, and Andrea Ninfo. ”Morphological analysis for architectural applications: Comparison between laser scan- ning and structure-from-motion photogrammetry.” Journal of Surveying Engineering 142.3 (2016): 04016004

    work page 2016

  2. [2]

    ”3D Gaussian Splatting for Real-Time Radiance Field Rendering.” ACM Transactions on Graphics (TOG) 42.4 (2023): 139-1

    Kerbl, Bernhard, et al. ”3D Gaussian Splatting for Real-Time Radiance Field Rendering.” ACM Transactions on Graphics (TOG) 42.4 (2023): 139-1

    work page 2023

  3. [3]

    ”NeRF–: Neural radiance fields without known camera parameters.” arXiv preprint arXiv:2102.07064 (2021)

    Wang, Zirui, et al. ”NeRF–: Neural radiance fields without known camera parameters.” arXiv preprint arXiv:2102.07064 (2021)

    arXiv 2021

  4. [4]

    ”Deepstereo: Learning to predict new views from the world’s imagery.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2016)

    Flynn, John, et al. ”Deepstereo: Learning to predict new views from the world’s imagery.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2016)

    work page 2016

  5. [5]

    Srinivasan, Matthew Tancik, Jonathan T

    Mildenhall, Ben, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. ”NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis.” Proceedings of the European Conference on Computer Vision (ECCV) , 2020

    work page 2020

  6. [6]

    Kulkarni, Bharath Ravindran, and Joshua B

    Li, Zhoutong, Tejas D. Kulkarni, Bharath Ravindran, and Joshua B. Tenenbaum. ”Learning Neural 3D Video Synthesis from Multi-View Video.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2021

    work page 2021

  7. [7]

    ”Neural Sparse V oxel Fields.” Proceedings of the 2020 ACM SIGGRAPH/Eurographics Symposium on Computer Animation , 2020

    Liu, Lingjie, Jiatao Gu, Zawlin Chen, Christian Theobalt, and Taku Komura. ”Neural Sparse V oxel Fields.” Proceedings of the 2020 ACM SIGGRAPH/Eurographics Symposium on Computer Animation , 2020

    work page 2020

  8. [8]

    ”GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020

    Niemeyer, Michael, Lars Mescheder, Michael Oechsle, and Andreas Geiger. ”GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020

    work page 2020

Show all 16 references
  1. [9]

    ”Citing Information: APA Style.” UNC Library Research Guides , accessed July 11, 2024

    University of North Carolina at Chapel Hill. ”Citing Information: APA Style.” UNC Library Research Guides , accessed July 11, 2024. https://guides.lib.unc.edu/c.php?g=716413&p=5098040

    2024

  2. [10]

    ”Gaussian Splatting GitHub Repository.” GitHub, accessed July 11, 2024

    GraphDeco-Inria. ”Gaussian Splatting GitHub Repository.” GitHub, accessed July 11, 2024. https://github.com/graphdeco-inria/gaussian- splatting

    2024

  3. [11]

    ”Gaussian Splatting: 3D Reconstruction and Novel View Synthesis, a Review.” IEEE Access (2024)

    Dalal, Anurag, et al. ”Gaussian Splatting: 3D Reconstruction and Novel View Synthesis, a Review.” IEEE Access (2024)

    2024

  4. [12]

    ”Camorph GitHub Repository.” GitHub, accessed July 11, 2024

    Fraunhofer-IIS. ”Camorph GitHub Repository.” GitHub, accessed July 11, 2024. https://github.com/Fraunhofer-IIS/camorph

    2024

  5. [13]

    ”3D Gaussian Splatting for UA V night photogrammetry.” International Symposium on Applied Geoinformatics (ISAG2024), 9-10 May 2024, Wrocław, Poland

    Burdziakowski, Pawel, et al. ”3D Gaussian Splatting for UA V night photogrammetry.” International Symposium on Applied Geoinformatics (ISAG2024), 9-10 May 2024, Wrocław, Poland. International Society for Photogrammetry and Remote Sensing (ISPRS), 2024

    2024

  6. [14]

    ”A survey on 3d gaussian splatting.” arXiv preprint arXiv:2401.03890 (2024)

    Chen, Guikun, and Wenguan Wang. ”A survey on 3d gaussian splatting.” arXiv preprint arXiv:2401.03890 (2024)

    2024 arXiv

  7. [15]

    Lerma, and Carmen P ´erez

    Portal ´es, Cristina, Jos ´e L. Lerma, and Carmen P ´erez. ”Photogrammetry and augmented reality for cultural heritage applications.” The Pho- togrammetric Record 24.128 (2009): 316-331

    2009

  8. [16]

    ”Evaluating Road Lighting Effects on Traffic Safety around Critical Road Sites.” 17th COTA International Conference of Transportation Professionals

    Sun, Cuicui, Yueru Xu, and Zhirui Ye. ”Evaluating Road Lighting Effects on Traffic Safety around Critical Road Sites.” 17th COTA International Conference of Transportation Professionals . Reston, V A: American Society of Civil Engineers, 2017. 7

    2017

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.