Pith. sign in

REVIEW 1 major objections 3 minor 55 references

A single jointly optimized 3D Gaussian Splatting model reconstructs RGB, red-edge, and NIR bands better than separate per-band models.

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

2026-08-05 12:59 UTC pith:E7P7MQ33

load-bearing objection Useful empirical recipe for multi-spectral 3DGS, but the headline superiority claim of JOINT-OPTIMIZED is not statistically supported by the per-scene data. the 1 major comments →

arxiv 2509.00989 v1 pith:E7P7MQ33 submitted 2025-08-31 cs.CV

Towards Integrating Multi-Spectral Imaging with Gaussian Splatting

Josef Gr\"un , Lukas Meyer , Maximilian Weiherer , Bernhard Egger , Marc Stamminger , Linus Franke This is my paper
classification cs.CV
keywords multi-spectral imaging3D Gaussian splattingspectral cross-talkjoint optimizationnovel view synthesismulti-spectral densificationspherical harmonicsagricultural phenotyping
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.

Multi-spectral cameras add bands beyond visible light, but those narrow-band images are too weak to drive a 3D Gaussian Splatting model on their own. The paper compares three ways to put RGB plus red, green, red-edge, and near-infrared images into 3DGS: fully separate per-band models, splitting off from a finished RGB model, and joint optimization of one model whose Gaussians hold colors for every band. It argues that the joint strategy, with a delayed RGB-only start and a multi-spectral-aware densification step, gives the best reconstruction in every band and even improves RGB by letting the extra bands feed geometry back into the visible channels. If correct, the practical recipe is to integrate spectral data directly into the Gaussian color representation and share one geometry across all wavelengths, rather than treating each band as an independent scene.

Core claim

On its own terms, the paper's central claim is that a single 3DGS model with shared Gaussian geometry and per-band spherical-harmonic colors is the right way to fuse multi-spectral imagery, provided the training schedule is staged. The authors define three strategies: SEPARATE (one model per band), SPLIT (RGB model first, then copy geometry and fit each band), and JOINT (all bands optimized together, with a random band rendered per iteration). They report that JOINT with two modifications—SpecDelay, which keeps the first 30,000 iterations RGB-only, and MSAD, which densifies when any band's gradient exceeds threshold—achieves the best aggregate PSNR/SSIM/LPIPS over all bands (24.99 PSNR vs 23

What carries the argument

The central object is a joint 3DGS model in which each Gaussian carries a full set of spherical-harmonic color parameters—one set per spectral band (seven total)—while position, scale, rotation, and opacity are shared across all bands. Each training iteration randomly selects one band and renders only that band, so gradients from all modalities accumulate on the same structure. The two load-bearing mechanisms are the spectral delay (RGB-only warm start so the dense, feature-rich RGB data builds trustworthy geometry before noisy narrow-band channels arrive) and multi-spectral-aware densification (MSAD), which splits or clones a Gaussian when any single band's accumulated gradient crosses the

Load-bearing premise

The load-bearing premise is that all spectral bands see exactly the same scene geometry and are registered to one common coordinate frame, so a single shared set of Gaussian positions, scales, and opacities can be optimized for all bands at once; if cameras are misaligned or geometry truly differs by wavelength, the joint strategy's cross-talk benefit collapses.

What would settle it

Train JOINT-OPTIMIZED and the SPLIT baseline on a scene where the NIR or red-edge camera has a deliberately shifted pose (a few centimeters or a couple of pixels in the validation set) while RGB stays fixed, then compare per-band PSNR. If the joint model's all-band score drops below SPLIT under realistic misalignment, the shared-geometry assumption is the exact load-bearing component; if it tolerates the shift, the recipe is more robust than its stated assumption.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Separate per-band 3DGS models are the weakest option; sharing one geometry across bands raises every band's reconstruction quality.
  • A joint model with SpecDelay and MSAD beats both the naive joint model and the RGB-then-split pipeline on PSNR, SSIM, and LPIPS across the seven bands.
  • Adding spectral bands improves RGB reconstruction too: all-band joint training gains about 0.2 dB over RGB-only, with the largest RGB gain from combining green, red, and red-edge.
  • The joint representation is more compact: roughly two million Gaussians cover a scene that SEPARATE reconstructs with about one million splats per band.
  • The training recipe—RGB delay, spectral warm-up, extended densification—is the transferable guidance for any multi-spectral 3DGS deployment.

Where Pith is reading between the lines

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

  • The paper's success depends on pixel-perfect alignment across cameras; an untested stress test would be to inject small pose offsets or parallax between RGB and NIR and see when the joint strategy's advantage over SPLIT disappears.
  • Because the improvements are attributed to shared geometry, the same recipe may transfer to other modalities that share geometry, such as RGB plus thermal in scenes without strong emissive differences, or RGB plus depth—but the paper's Section 5 limitation warns that thermal geometry can genuinely differ.
  • The cross-talk here is geometric only; coupling this training schedule with a neural or feature-based color representation could extend the benefit to color-level cross-talk, which the paper flags as a limitation of its JOINT-OPTIMIZED configuration.
  • For agricultural use, the practical payoff would be spatially consistent vegetation indices computed directly from the joint radiance field rather than from separately aligned band maps; the paper suggests fruit counting and phenotyping but does not test them.

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

1 major / 3 minor

Summary. The paper investigates how to integrate RGB and multi-spectral imagery (red, green, red-edge, NIR) into 3D Gaussian Splatting. It formulates and compares three strategies—SEPARATE (independent per-band 3DGS models), SPLIT (RGB-first geometry copy then per-band refinement), and JOINT (single model with per-Gaussian multi-spectral SH coefficients)—and proposes training modifications (SpecDelay, ExtADC, MSAD, SIG). The main claim is that a staged joint optimization, labeled JOINT-OPTIMIZED, yields the best overall spectral reconstruction and improves RGB via spectral cross-talk. Experiments are run on the authors' seven-scene drone multi-spectral dataset, with PSNR/SSIM/LPIPS reported as scene averages. The paper also provides an ablation study of the proposed modules and a cross-talk analysis that systematically adds spectral bands to RGB.

Significance. If the central claim held, the paper would offer a practical, interpretable recipe for multi-spectral 3DGS: representing each band through direct SH coefficients and sharing geometry across bands, with an RGB warm-up and spectrally aware densification. The study is systematic (three strategies, four modules, seven scenes), the code is promised open-source, and the per-scene tables in the supplement are a transparency plus. However, the headline conclusions rest almost entirely on averaged metrics over seven scenes, and the supplementary per-scene numbers do not show a consistent advantage for the proposed method. The RGB cross-talk improvement is small (~0.2 dB) and not statistically characterized. Because the evaluation is the only evidence for the core claims, the paper's significance is currently contingent on substantially stronger statistical and comparative support.

major comments (1)
  1. [Sec. 3.1, Sec. 3.2.4, Sec. 2.2] The dataset, globally aligned point cloud, poses, and the MSAD densification module all come from the authors' prior work [MGW*25], yet that method is not included as a quantitative baseline. SPLIT and JOINT are compared only against plain 3DGS per band. Since [MGW*25] is the most relevant multi-spectral 3DGS baseline and provides the input representation and a core module, omitting it makes the 'state of the art' positioning unclear and leaves open the possibility that the proposed gains are not competitive with the existing neural-color approach. Adding this baseline is necessary to support the claim that direct SH representation is preferable.
minor comments (3)
  1. [Sec. 1, Sec. 3.2.1] The text uses 'naïve' and 'SEPARATE' inconsistently (e.g., 'Naively' in the introduction and 'naïve' later). Please unify spelling and capitalization of strategy names throughout, including table captions.
  2. [Fig. 2] The caption 'S����� ��J����-O��������' is corrupted/misrendered. Please fix the caption text and ensure the figure labels match the main text.
  3. [Sec. 4.4 and Table S.T8/S.T9] The primitive-count claim states that MSAD increases splats by ~100% compared to plain JOINT. In the GARDEN table, MSAD variants use 1.9–2.2M vs ~0.95M for JOINT, which is consistent, but in LAKE the increase is from 2.31M to 4.86M, also consistent. However, the text does not discuss why the GARDEN and LAKE counts differ so drastically; a sentence on scene-scale dependence would help.

Circularity Check

0 steps flagged

No circularity: the JOINT-vs-SEPARATE/SPLIT comparison is empirical and self-contained; same-author reuse affects provenance, not the derivation.

full rationale

The paper's central claims are comparative measurements, not derived predictions. JOINT, SPLIT, and SEPARATE are distinct training strategies evaluated on held-out validation views with standard metrics; no equation defines one strategy's outcome in terms of another, and no fitted parameter is relabeled as a prediction. The reuse of the authors' prior work [MGW*25] supplies the dataset, the initial aligned point cloud, and the MSAD densification heuristic, but the superiority of JOINT-OPTIMIZED over SEPARATE/SPLIT is established by the experiments in Tables 1 and 2, not by the citation itself. The paper also does not invoke a uniqueness theorem or smuggle in an ansatz via citation. The Limitations section explicitly flags the shared-geometry assumption and the geometric-only nature of cross-talk, which are honest scope restrictions rather than circular steps. Per-scene variability and the absence of significance tests are statistical concerns, not circularity: the averages could be non-representative, but that does not make the derivation tautological. Under the stated criteria, no quoted equation or definition reduces to its own input, so the appropriate score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The method adds per-band spherical harmonics color parameters and training-schedule modifications, not new particles, forces, or dimensions. The free parameters are hand-chosen training hyperparameters rather than fitted physical constants. The main assumptions are the shared-geometry premise and reliance on the authors' prior alignment and dataset pipeline.

free parameters (4)
  • SpecDelay duration = 30,000 iterations
    Hand-chosen training schedule; RGB-only for first 30k iterations (Sec 3.2.4).
  • ExtADC densification interval = 60,000 iterations
    Hand-chosen extension of the densification period (Sec 3.2.4).
  • Multi-spectral warm-up window = 3,000 iterations between 29k and 32k
    Hand-chosen to stabilize spectral introduction (Sec 3.3.3).
  • Learning rate = 0.005
    Standard 3DGS hyperparameter (Sec 3.3.1).
axioms (4)
  • domain assumption A single 3D Gaussian Splatting model can represent all spectral bands with shared geometry and per-band spherical harmonics colors.
    Central modeling choice in Sec 3.2.3; validated only on VIS/NIR data.
  • domain assumption A globally aligned sparse point cloud and per-camera poses are available for all spectral cameras.
    Stated in Sec 3.1; supplied by the authors' prior pipeline [MGW*25].
  • standard math The 3DGS differentiable rendering and photometric loss (Eq. 1) are a valid objective for multi-spectral reconstruction.
    Borrowed from Kerbl et al. [KKLD23]; used without modification.
  • domain assumption Ground-truth images in the evaluation dataset are correctly aligned for metric comparison.
    The dataset is from the authors' own prior work (Sec 4), so alignment and ground truth are self-produced.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 24801 in / 14562 out tokens · 149730 ms · 2026-08-05T12:59:18.528846+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Towards Integrating Multi-Spectral Imaging with Gaussian Splatting." pith.science (2026). https://pith.science/paper/E7P7MQ33

@misc{pith2026250900989,
  author       = {Pith},
  title        = {Pith review of: Towards Integrating Multi-Spectral Imaging with Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7P7MQ33}},
  note         = {Machine review of arXiv:2509.00989}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present a study of how to integrate color (RGB) and multi-spectral imagery (red, green, red-edge, and near-infrared) into the 3D Gaussian Splatting (3DGS) framework, a state-of-the-art explicit radiance-field-based method for fast and high-fidelity 3D reconstruction from multi-view images. While 3DGS excels on RGB data, naive per-band optimization of additional spectra yields poor reconstructions due to inconsistently appearing geometry in the spectral domain. This problem is prominent, even though the actual geometry is the same, regardless of spectral modality. To investigate this, we evaluate three strategies: 1) Separate per-band reconstruction with no shared structure. 2) Splitting optimization, in which we first optimize RGB geometry, copy it, and then fit each new band to the model by optimizing both geometry and band representation. 3) Joint, in which the modalities are jointly optimized, optionally with an initial RGB-only phase. We showcase through quantitative metrics and qualitative novel-view renderings on multi-spectral datasets the effectiveness of our dedicated optimized Joint strategy, increasing overall spectral reconstruction as well as enhancing RGB results through spectral cross-talk. We therefore suggest integrating multi-spectral data directly into the spherical harmonics color components to compactly model each Gaussian's multi-spectral reflectance. Moreover, our analysis reveals several key trade-offs in when and how to introduce spectral bands during optimization, offering practical insights for robust multi-modal 3DGS reconstruction.

Figures

Figures reproduced from arXiv: 2509.00989 by Bernhard Egger, Josef Gr\"un, Linus Franke, Lukas Meyer, Marc Stamminger, Maximilian Weiherer.

Figure 1
Figure 1. Figure 1: We explore the integration of multi-spectral captures containing RGB and further spectral bands (e.g., red edge, near infra-red) into the 3D Gaussian Splatting (3DGS) framework. We investigate three main strategies: SEPARATE individually optimizes a 3DGS model for each band, whereas SPLIT initially constructs an RGB model, replicates the geometry, and then fits each channel separately through geometry and … view at source ↗
Figure 2
Figure 2. Figure 2: A visual comparison of the NIR reconstructions using the SEPARATE strategy and the JOINT-OPTIMIZED configuration is shown. The SEPARATE strategy significantly lacks geometric detail, whereas the same region in the jointly optimized reconstruction exhibits improved geometry as a result of the cross-talk effect [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: for a schematic visualization of SpecDelay. The objective of this training configuration is to evaluate whether postponing spectral optimization and leveraging the strong geometric reconstruction of RGB data improves the reconstruction quality of the spectral bands. Extended ADC (ExtADC) Additionally, we propose an extended ADC configuration, where the densification interval is increased to 60,000 iteratio… view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison on the SINGLE TREE scene with all available spectral bands. 5. Limitations In the preceding section, we demonstrated the effect of the JOINT￾OPTIMIZED configuration on the quantitative improvements, which can be attributed to spectral cross-talk. A major limitation, how￾ever, is that only geometric cross-talk is possible, as the JOINT￾OPTIMIZED configuration allows for joint optimization … 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

55 extracted references · 48 canonical work pages · 3 internal anchors

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry.original add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 i...

  3. [3]

    : Multispectral and hyperspectral remote sensing for identification and mapping of wetland vegetation: a review

    Adam E., Mutanga O., Rugege D. : Multispectral and hyperspectral remote sensing for identification and mapping of wetland vegetation: a review. Wetlands Ecology and Management 18, 3 (jun 2010), 281--296

    work page 2010

  4. [4]

    : Neural point-based graphics

    Aliev K.-A., Sevastopolsky A., Kolos M., Ulyanov D., Lempitsky V. : Neural point-based graphics. In ECCV (2020), pp. 696--712

    work page 2020

  5. [5]

    T., Knoll P., Li Y.-H., Barthel F., Hilsmann A., Eisert P., Morgenstern W

    Bagdasarian M. T., Knoll P., Li Y.-H., Barthel F., Hilsmann A., Eisert P., Morgenstern W. : 3dgs.zip: A survey on 3d gaussian splatting compression methods, 2024

    work page 2024

  6. [6]

    T., Mildenhall B., Verbin D., Srinivasan P

    Barron J. T., Mildenhall B., Verbin D., Srinivasan P. P., Hedman P. : Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR (2022), pp. 5470--5479

    work page 2022

  7. [7]

    T., Mildenhall B., Verbin D., Srinivasan P

    Barron J. T., Mildenhall B., Verbin D., Srinivasan P. P., Hedman P. : Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV (October 2023), pp. 19640--19648

    work page 2023

  8. [8]

    K., Ottou T

    Chen G., Narayanan S. K., Ottou T. G., Missaoui B., Muriki H., Pradalier C., Chen Y. : Hyperspectral neural radiance fields. arXiv preprint arXiv:2403.14839 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    : Thermal3d-gs: Physics-induced 3d gaussians for thermal infrared novel-view synthesis

    Chen Q., Shu s., Bai X. : Thermal3d-gs: Physics-induced 3d gaussians for thermal infrared novel-view synthesis. In ECCV (2024)

    work page 2024

  10. [10]

    https://ag.dji.com/mavic-3-m, 2025

    DJI : Mavic 3 multispectral (3m). https://ag.dji.com/mavic-3-m, 2025. Accessed: 2025-01-24

    work page 2025

  11. [11]

    : Vr-splatting: Foveated radiance field rendering via 3d gaussian splatting and neural points

    Franke L., Fink L., Stamminger M. : Vr-splatting: Foveated radiance field rendering via 3d gaussian splatting and neural points. Proc. ACM Comput. Graph. Interact. Tech. 8, 1 (May 2025)

    work page 2025

  12. [12]

    : Livenvs: Neural view synthesis on live rgb-d streams

    Fink L., R \"u ckert D., Franke L., Keinert J., Stamminger M. : Livenvs: Neural view synthesis on live rgb-d streams. In ACM SIGGRAPH Asia 2023 Conference Proceedings (2023)

    work page 2023

  13. [13]

    : Vet: Visual error tomography for point cloud completion and high-quality neural rendering

    Franke L., R\" u ckert D., Fink L., Innmann M., Stamminger M. : Vet: Visual error tomography for point cloud completion and high-quality neural rendering. In SIGGRAPH Asia 2023 Conference Papers (2023), SA '23, Association for Computing Machinery

    work page 2023

  14. [14]

    : Trips: Trilinear point splatting for real-time radiance field rendering

    Franke L., R \"u ckert D., Fink L., Stamminger M. : Trips: Trilinear point splatting for real-time radiance field rendering. Computer Graphics Forum 43, 2 (2024)

    work page 2024

  15. [15]

    : Inovis: Instant novel-view synthesis

    Harrer M., Franke L., Fink L., Stamminger M., Weyrich T. : Inovis: Instant novel-view synthesis. In SIGGRAPH Asia Conference Papers (2023), Association for Computing Machinery

    work page 2023

  16. [16]

    : Thermonerf: Multimodal neural radiance fields for thermal novel view synthesis

    Hassan M., Forest F., Fink O., Mielle M. : Thermonerf: Multimodal neural radiance fields for thermal novel view synthesis. arXiv 2403.12154 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    : INPC : Implicit neural point clouds for radiance field rendering

    Hahlbohm F., Franke L., Kappel M., Castillo S., Eisemann M., Stamminger M., Magnor M. : INPC : Implicit neural point clouds for radiance field rendering. In 3DV (2025)

    work page 2025

  18. [18]

    : Efficient perspective-correct 3d gaussian splatting using hybrid transparency

    Hahlbohm F., Friederichs F., Weyrich T., Franke L., Kappel M., Castillo S., Stamminger M., Eisemann M., Magnor M. : Efficient perspective-correct 3d gaussian splatting using hybrid transparency. Computer Graphics Forum 44, 2 (2025)

    work page 2025

  19. [19]

    : Phenotypic techniques and applications in fruit trees: a review

    Huang Y., Ren Z., Li D., Liu X. : Phenotypic techniques and applications in fruit trees: a review. Plant Methods 16, 1 (2020)

    work page 2020

  20. [20]

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

    Kerbl B., Kopanas G., Leimk \"u hler T., Drettakis G. : 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42, 4 (July 2023)

    work page 2023

  21. [21]

    : A hierarchical 3d gaussian representation for real-time rendering of very large datasets

    Kerbl B., Meuleman A., Kopanas G., Wimmer M., Lanvin A., Drettakis G. : A hierarchical 3d gaussian representation for real-time rendering of very large datasets. TOG 43, 4 (2024), 1--15

    work page 2024

  22. [22]

    : Thermalgaussian: Thermal 3d gaussian splatting, 2024

    Lu R., Chen H., Zhu Z., Qin Y., Lu M., Zhang L., Yan C., Xue A. : Thermalgaussian: Thermal 3d gaussian splatting, 2024

    work page 2024

  23. [23]

    : Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis

    Luiten J., Kopanas G., Leibe B., Ramanan D. : Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV (2024)

    work page 2024

  24. [24]

    : Spectralnerf: physically based spectral rendering with neural radiance field

    Li R., Liu J., Liu G., Zhang S., Zeng B., Liu S. : Spectralnerf: physically based spectral rendering with neural radiance field. In AAAI'24/IAAI'24/EAAI'24 (2024), AAAI Press

    work page 2024

  25. [25]

    : Spec-nerf: Multi-spectral neural radiance fields

    Li J., Li Y., Sun C., Wang C., Xiang J. : Spec-nerf: Multi-spectral neural radiance fields. In IEEE International Conference on Acoustics, Speech and Signal Processing (2024), pp. 2485--2489

    work page 2024

  26. [26]

    : Vastgaussian: Vast 3d gaussians for large scene reconstruction

    Lin J., Li Z., Tang X., Liu J., Liu S., Liu J., Lu Y., Wu X., Xu S., Yan Y., Yang W. : Vastgaussian: Vast 3d gaussians for large scene reconstruction. In CVPR (2024)

    work page 2024

  27. [27]

    Y., Pan X.-Y., Fridovich-Keil S., Wetzstein G

    Lin Y. Y., Pan X.-Y., Fridovich-Keil S., Wetzstein G. : ThermalNeRF : Thermal radiance fields. In IEEE International Conference on Computational Photography (ICCP) (2024), IEEE

    work page 2024

  28. [28]

    FruitNeRF++: A Generalized Multi-Fruit Counting Method Utilizing Contrastive Learning and Neural Radiance Fields

    Meyer L., Ardelean A.-T., Weyrich T., Stamminger M. : Fruitnerf++: A generalized multi-fruit counting method utilizing contrastive learning and neural radiance fields. arXiv2505.19863 2505.19863 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  29. [29]

    : Instant neural graphics primitives with a multiresolution hash encoding

    M\" u ller T., Evans A., Schied C., Keller A. : Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (TOG) 41, 4 (jul 2022)

    work page 2022

  30. [30]

    : Fruitnerf: A unified neural radiance field based fruit counting framework

    Meyer L., Gilson A., Schmid U., Stamminger M. : Fruitnerf: A unified neural radiance field based fruit counting framework. In IROS (2024)

    work page 2024

  31. [31]

    : Multi-spectral gaussian splatting with neural color representation, 2025

    Meyer L., Grün J., Weiherer M., Egger B., Stamminger M., Franke L. : Multi-spectral gaussian splatting with neural color representation, 2025

    work page 2025

  32. [32]

    Mildenhall B., Srinivasan P. P. ., Tancik M., Barron Jonathan T.and Ramamoorthi R., Ng R. : Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV (2020), pp. 405--421

    work page 2020

  33. [33]

    : Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with 900+ fps

    Niemeyer M., Manhardt F., Rakotosaona M.-J., Oechsle M., Duckworth D., Gosula R., Tateno K., Bates J., Kaeser D., Tombari F. : Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with 900+ fps. arXiv preprint arXiv:2403.13806 (2024)

    Pith/arXiv arXiv 2024

  34. [34]

    T., Petersson L

    Namin S. T., Petersson L. : Classification of materials in natural scenes using multi-spectral images. In International Conference on Intelligent Robots and Systems (2012), IEEE, pp. 1393--1398

    work page 2012

  35. [35]

    : Compressed 3d gaussian splatting for accelerated novel view synthesis

    Niedermayr S., Stumpfegger J., Westermann R. : Compressed 3d gaussian splatting for accelerated novel view synthesis. In CVPR (June 2024), pp. 10349--10358

    work page 2024

  36. [36]

    : Exploring multi-modal neural scene representations with applications on thermal imaging

    \"Ozer M., Weiherer M., Hundhausen M., Egger B. : Exploring multi-modal neural scene representations with applications on thermal imaging. ArXiv 2403.11865 (2024)

    Pith/arXiv arXiv 2024

  37. [37]

    : Cross-spectral neural radiance fields

    Poggi M., Zama Ramirez P., Tosi F., Salti S., Di Stefano L., Mattoccia S. : Cross-spectral neural radiance fields. In 3DV (2022). 3DV

    work page 2022

  38. [38]

    : StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering

    Radl L., Steiner M., Parger M., Weinrauch A., Kerbl B., Steinberger M. : StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering . ACM TOG 4, 43 (2024)

    work page 2024

  39. [39]

    Shum H.-Y., Chan S.-C., Kang S. B. : Image-based rendering. Springer Science & Business Media, 2008

    work page 2008

  40. [40]

    L., Frahm J.-M

    Sch\" o nberger J. L., Frahm J.-M. : Structure-from-motion revisited. In CVPR (2016), pp. 4104--4113

    work page 2016

  41. [41]

    N., Graf H., Weinmann M

    Sinha S. N., Graf H., Weinmann M. : Spectralgaussians: Semantic, spectral 3d gaussian splatting for multi-spectral scene representation, visualization and analysis, 2024

    work page 2024

  42. [42]

    Shum H., Kang S. B. : Review of image-based rendering techniques. In Visual Communications and Image Processing 2000 (2000), vol. 4067, SPIE, pp. 2--13

    work page 2000

  43. [43]

    L., Zheng E., Pollefeys M., Frahm J.-M

    Sch\" o nberger J. L., Zheng E., Pollefeys M., Frahm J.-M. : Pixelwise view selection for unstructured multi-view stereo. In ECCV (2016)

    work page 2016

  44. [44]

    B., Zollhöfer M

    Tewari A., Fried O., Thies J., Sitzmann V., Lombardi S., Sunkavalli K., Martin-Brualla R., Simon T., Saragih J., Nießner M., Pandey R., Fanello S., Wetzstein G., Zhu J.-Y., Theobalt C., Agrawala M., Shechtman E., Goldman D. B., Zollhöfer M. : State of the art on neural rendering. CGF 39, 2 (2020), 701--727

    work page 2020

  45. [45]

    HyperGS: Hyperspectral 3D Gaussian Splatting

    Thirgood C., Mendez O., Ling E., Storey J., Hadfield S. : Hypergs: Hyperspectral 3d gaussian splatting. arXiv 2412.12849 (12 2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  46. [46]

    T., Wetzstein G., Zollhoefer M., Golyanik V

    Tewari A., Thies J., Mildenhall B., Srinivasan P., Tretschk E., Wang Y., Lassner C., Sitzmann V., Martin-Brualla R., Lombardi S., Simon T., Theobalt C., Niessner M., Barron J. T., Wetzstein G., Zollhoefer M., Golyanik V. : Advances in neural rendering. CGF 41, 2 (2022), 703--735

    work page 2022

  47. [47]

    : Nerfstudio: A modular framework for neural radiance field development

    Tancik M., Weber E., Ng E., Li R., Yi B., Wang T., Kristoffersen A., Austin J., Salahi K., Ahuja A., Mcallister D., Kerr J., Kanazawa A. : Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings (2023)

    work page 2023

  48. [48]

    : 4d gaussian splatting for real-time dynamic scene rendering

    Wu G., Yi T., Fang J., Xie L., Zhang X., Wei W., Liu W., Tian Q., Wang X. : 4d gaussian splatting for real-time dynamic scene rendering. In CVPR (2024), pp. 20310--20320

    work page 2024

  49. [49]

    P., Patel V

    Xu J., Liao M., Kathirvel R. P., Patel V. M. : Leveraging thermal modality to enhance reconstruction in low-light conditions. In ECCV (2024), Leonardis A., Ricci E., Roth S., Russakovsky O., Sattler T., Varol G., (Eds.)

    work page 2024

  50. [50]

    : Mip-splatting: Alias-free 3d gaussian splatting

    Yu Z., Chen A., Huang B., Sattler T., Geiger A. : Mip-splatting: Alias-free 3d gaussian splatting. In CVPR (2024), pp. 19447--19456

    work page 2024

  51. [51]

    : Learning a locally unified 3d point cloud for view synthesis

    You M., Guo M., Lyu X., Liu H., Hou J. : Learning a locally unified 3d point cloud for view synthesis. IEEE Transactions on Image Processing 32 (2023), 5610--5622

    work page 2023

  52. [52]

    : gsplat: An open-source library for gaussian splatting

    Ye V., Li R., Kerr J., Turkulainen M., Yi B., Pan Z., Seiskari O., Ye J., Hu J., Tancik M., Kanazawa A. : gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research 26, 34 (2025), 1--17

    work page 2025

  53. [53]

    Thermal-NeRF: Neural Radiance Fields from an Infrared Camera

    Ye T., Wu Q., Deng J., Liu G., Liu L., Xia S., Pang L., Yu W., Pei L. : Thermal-nerf: Neural radiance fields from an infrared camera. arXiv 2403.10340 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  54. [54]

    : Surface splatting

    Zwicker M., Pfister H., Van Baar J., Gross M. : Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques (2001), pp. 371--378

    work page 2001

  55. [55]

    : Nerf++: Analyzing and improving neural radiance fields

    Zhang K., Riegler G., Snavely N., Koltun V. : Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492 (2020)

    Pith/arXiv arXiv 2010