Pith. sign in

REVIEW 4 major objections 5 minor 48 references

UnMix-NeRF: Spectral Unmixing Meets Neural Radiance Fields

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read UnMix-NeRF embeds spectral unmixing into a neural radiance field so that one jointly optimized model produces hyperspectral novel views, unsupervised material segmentations, and editable scene appearance, outperforming prior…

desk verdict Spectral view synthesis is solid and worth publishing; the unsupervised material segmentation claim is circular and unsupported as written. read the letter →

arxiv 2506.21884 v2 pith:35REWCQD submitted 2025-06-27 eess.IV cs.AIcs.CVcs.LGeess.SP

Fabian Perez , Sara Rojas , Carlos Hinojosa , Hoover Rueda-Chacón , Bernard Ghanem This is my paper · ORCID
classification eess.IVcs.AIcs.CVcs.LGeess.SP
keywords neuralradiancefieldsspectralunmixinghyperspectralimagingnovelviewsynthesisunsupervisedmaterialsegmentationendmemberdictionarysceneeditingvolumetricrendering
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

UnMix-NeRF claims that a neural radiance field—a network that maps 3D points to density and color and renders images by integrating along rays—can be built on spectral unmixing rather than on per-wavelength color regression. Giving the network a shared dictionary of global material spectra (endmembers) and per-point mixture weights (abundances) makes the same optimization that reconstructs hyperspectral novel views also yield an unsupervised segmentation of the scene into materials and support editing by swapping dictionary entries. The paper reports that this formulation improves hyperspectral reconstruction over existing NeRF and Gaussian-splatting baselines on three datasets while cutting training time compared with NeSpoF. If the factorization holds, material-level scene understanding for robotics, augmented reality, and simulation comes out of the same field that renders the images.

What carries the argument

The load-bearing object is a global endmember dictionary $\mathbf{E}\in\mathbb{R}^{B\times K}$, whose columns are pure material spectral signatures learned jointly with the network and clamped to $[0,1]$. At each point the field outputs abundances $\tilde{\mathbf{a}}$ (softmax, so non-negative and sum-to-one) and scaling factors $\tilde{\mathbf{s}}$ (sigmoid), and the diffuse reflectance is $\mathbf{E}\tilde{\mathbf{S}}\tilde{\mathbf{a}}$; a specular head predicts $\mathbf{c}_s$ modulated by a tint factor $h$. This dictionary does the real work: rendered signatures are compared against it to assign material clusters, and scene editing means replacing or modifying its columns. The paper's central move is betting that the unsupervised dictionary plus abundances is enough to link reconstruction quality to physical material identity.

What would settle it

Take a scene containing K known materials, measure their true spectra separately in the lab, train UnMix-NeRF, and compare each learned endmember column with the measured spectra by spectral angle while also checking whether abundance maps align with the known spatial material layout. The central claim would be falsified if reconstruction PSNR remains high but the best endmember-to-material spectral angles are large and the abundance maps do not track the known material regions.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the Extended Linear Mixing Model can be placed inside the NeRF volume renderer. Every 3D point predicts an abundance vector $\mathbf{a}$ and per-material scaling factors $\mathbf{s}$ over a global learnable endmember matrix $\mathbf{E}$, forming the diffuse reflectance $\mathbf{c}_d = \mathbf{E}\tilde{\mathbf{S}}\tilde{\mathbf{a}}$; a view-dependent branch adds a tinted specular term $h\mathbf{c}_s$ according to the dichromatic model. The per-point abundances are volume-rendered along rays to produce abundance maps, and material segmentation is obtained by assigning each rendered spectral signature to the closest endmember through a normalized inner-product softmax. The paper argues that optimizing spectral and RGB reconstruction losses alone is sufficient to guide the unmixing, and reports that this formulation gives the best hyperspectral reconstruction among the compared baselines on NeSpoF, BaySpec, and Surface Optics scenes while also producing unsupervised material clusters.

Load-bearing premise

The load-bearing premise is that endmembers and abundances learned only from spectral and RGB reconstruction losses correspond to physically distinguishable materials in the scene, with no supervision or independent validation of that correspondence.

Editorial extensions

If this is right

  • Rendered abundance maps give unsupervised material labels as a by-product of hyperspectral novel view synthesis.
  • Editing the learned endmember dictionary changes one material's appearance everywhere in the scene, consistently across all views.
  • Hyperspectral novel view synthesis runs in roughly 44 minutes per NeSpoF scene, far below the 11.9-hour cost of the NeSpoF baseline.
  • Classical 2D unmixing initialization (VCA) stabilizes optimization and slightly improves reconstruction over random initialization.
  • No segmentation supervision is used; spectral and RGB reconstruction losses alone drive the reported material clustering.

Reading between the lines

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

  • A natural next step is to tie the learned dictionary to a known spectral library, regularizing each endmember toward a measured material spectrum; the present paper leaves the endmembers free, which is exactly where the physical-interpretation risk sits.
  • If the factorization transfers across scenes, endmembers could become a reusable material vocabulary, so a segmenter trained on one scene's abundances might label a different scene without retraining; the paper does not test cross-scene transfer.
  • Because segmentation relies on spectral distinguishability, spectrally similar but functionally different materials (metamers) should collapse into one cluster; an RGB-only version of the same architecture would test how much of the separation is due to spectral bands rather than geometry.
  • The hard cluster assignment via argmax discards the soft abundance information; using rendered abundance vectors directly as per-pixel mixture weights would give probabilistic material layers and might improve segmentation on fine-grained boundaries, which the paper does not explore.
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 UnMix-NeRF, a NeRF-based framework that integrates spectral unmixing (the extended linear mixing model, ELMM) into volumetric rendering for joint hyperspectral novel view synthesis and unsupervised material segmentation. A learnable global endmember dictionary E and per-point abundance/scaling predictions form a diffuse reflectance component, complemented by a view-dependent specular branch; the rendered spectral radiance is projected to RGB via a camera response matrix. Material segmentation is performed by a cluster probe that assigns each rendered spectral signature to the nearest learned endmember (Eq. 12–13). The method is evaluated on the synthetic NeSpoF dataset (with a newly generated material-label extension), the BaySpec dataset, and the Surface Optics dataset, reporting improved spectral reconstruction PSNR/SSIM/SAM/RMSE over several NeRF-, 3DGS-, and hyperspectral-specific baselines, alongside ablation studies, qualitative abundance maps, and an editing demonstration.

Significance. The spectral reconstruction results are strong and constitute the paper's clearest contribution: UnMix-NeRF consistently outperforms strong baselines such as HyperGS on BaySpec and Surface Optics, with substantially lower training time than NeSpoF (44 min vs 11.9 hours per scene), and the ablation in Table 1 supports the value of the scaling factors, RGB loss, and specular field. The proposed integration of spectral unmixing into a NeRF-style representation is a sensible and potentially useful direction for material-aware scene representation. However, the second headline claim—unsupervised material segmentation with 'pixel-accurate' performance—is not supported by the evidence presented: the segmentation probe of Eq. (12) directly re-encodes the learned dictionary, the reported F1=0.41 and mIoU=0.28 on the self-generated NeSpoF extension are modest, and there are no segmentation baselines, no K-sensitivity analysis, and no independent validation that the learned endmembers correspond to physical materials. The paper also contains an internal inconsistency between the text's claim of 'best overall' performance and Table 4.

major comments (4)
  1. [§4.4, Eq. (12)–(13)] The material segmentation probe uses the normalized inner product between the rendered full spectral radiance C(r) and the learned endmember matrix E. Since C(r) is built from E through Eq. (7) (plus the specular term), this assignment is a re-expression of the learned dictionary rather than an independent test of material decomposition. Moreover, contribution (ii) claims that segmentation leverages the rendered abundance vectors A(r), but Eq. (12) bypasses A(r) entirely and uses C(r). The authors should either use A(r) from Eq. (10) for segmentation, or justify why the C(r)-based probe is preferable, and must validate the resulting clusters against independent material ground truth rather than relying on the circular construction.
  2. [§5.4] The quantitative segmentation evaluation is insufficient to support the claim of 'superior material segmentation.' Only self-generated NeSpoF material maps are used as ground truth, with no comparison to any existing segmentation method (e.g., k-means clustering on rendered spectra or abundances, spectral angle-based labeling, SAM-based segmentation, or unsupervised baselines such as spectral clustering). The reported F1 of 0.41 and mIoU of 0.28 are low, and without baselines the reader cannot judge whether the proposed probe adds value over simple alternatives. Please add at least one straightforward spectral-clustering baseline and report per-scene segmentation results, along with variance across runs.
  3. [§4.1 and §5.1 (implementation details)] The number of endmembers K is selected as a 'heuristic approximation of the number of distinct materials,' making K a free parameter that directly controls the number of segmentation clusters and the editing capability. No sensitivity analysis over K is provided, and no criterion for choosing K is given. Since the segmentation and editing claims depend critically on K, the authors should include an ablation over K (e.g., K−1, K, K+1 relative to the heuristic) and, ideally, a discussion of when the learned dictionary is identifiable under the ELMM.
  4. [§4.5 and Table 4] The claim in §4.5 that minimizing Eq. (16) 'ensures physically consistent spectral unmixing' is not supported: the loss contains only reconstruction terms, with no regularization encouraging endmembers to match physical material spectra or abundances to be sparse or interpretable. Meanwhile, Table 4 shows that on the Rosemary scene HyperGS achieves higher SSIM (0.9845 vs 0.9355) than UnMix-NeRF, contradicting the text that 'Our method achieves the best performance in both the Rosemary and Basil scenes.' Please correct the text and discuss this trade-off (PSNR/spectral fidelity vs SSIM) rather than claiming superiority on all metrics.
minor comments (5)
  1. [§3.2] There is a typo: 'endmenmber' should be 'endmember' in the sentence defining E.
  2. [§5.1 (datasets)] The dataset extension is described only as 'generating material maps from the synthetic scenes'; please provide details of how these ground-truth material labels were created (e.g., from the scene's underlying geometry/material assignments), and make the extension publicly available or at least specify its release to allow reproducibility.
  3. [Eq. (12)] The denominator notation |E||C(r)| is ambiguous: E is a matrix and |·| is not defined. Please use explicit column norms or describe the normalization precisely.
  4. [§5.3 / Figure 5] The qualitative segmentation visualization shows only the Ajar scene; consider showing at least one more scene and comparing qualitatively with a baseline clustering to illustrate the method's behavior on specular and textured materials.
  5. [Tables 2–4] No variance or error bars are reported for the reconstruction metrics. Since the method has several stochastic components (random initialization, optimization), a small number of repeated runs with mean±std would strengthen the claims of superiority.

Circularity Check

1 steps flagged · score 6.0 of 10

The unsupervised material-segmentation probe reduces, by construction, to the same learned endmember dictionary used to synthesize the spectra; the novel-view-synthesis results are independent and non-circular.

  1. self definitional [Section 4.4, Eqs. (12)-(13); cf. Section 4.1, Eqs. (7)-(9)]
    "Given the predicted spectral signature for each ray C(r) ∈ R^B, we compute the normalized inner product against the dictionary of endmembers as p(r) = softmax(E⊤C(r) / |E||C(r)|). ... the diffuse reflectance is computed as follows: c_d = E S̃ ã. ... The final per-point spectral reflectance is obtained by combining the diffuse and specular components ... c = c_d + c̃_s."

    The segmentation probe is not an independent measurement of material identity. The rendered spectrum C(r) is the volumetrically accumulated sum of per-point radiances c = E S̃ ã + h c_s, so C(r) is itself a function of the same learned dictionary E. Therefore E⊤C(r), after normalization and softmax, is, up to the specular term, a deterministic re-expression of the learned abundances ã and scaling factors S̃ that were optimized to reconstruct the spectra. Assigning each ray to argmax_k p_k(r) returns a cluster structure already implied by the unmixing factorization, so it cannot independently confirm that the columns of E correspond to physical materials. Moreover, Eqs. (12)-(13) bypass the rendered abundance vectors A(r) that contribution (ii) claims are used for segmentation.

full rationale

The hyperspectral novel-view-synthesis contribution is self-contained and empirically grounded: Tables 2-4 compare against external baselines on NeSpoF, BaySpec, and Surface Optics with standard metrics, and Table 1 ablates each component. I find no circularity in that part of the derivation chain. The circularity is confined to the unsupervised material-segmentation claim. Section 4.4 defines segmentation as a normalized inner product between the rendered spectrum C(r) and the learned endmember matrix E, but C(r) is itself synthesized from E via Eqs. (7)-(9); hence the cluster assignment is a rearrangement of the fitted unmixing parameters rather than an independent test of material decomposition. The reported F1=0.41 and mIoU=0.28 on self-generated NeSpoF labels, with no segmentation baselines or ablations isolating the unmixing constraints, further weaken the claim, though these are validation weaknesses rather than additional circular steps. There is one minor self-citation ([34]) supporting the general premise that spectral signatures discriminate materials, but it is not load-bearing for the derivation. Overall, one central prediction reduces by construction, while the synthesis half remains independent, giving a score of 6.

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

The central segmentation claim depends on the unverified assumption that unsupervised spectral factorization yields real materials, plus several free parameters (K, loss weights, temperature) and the self-generated ground-truth labels. The reconstruction claim is more standard and relies on established volumetric rendering and linear mixing models.

free parameters (3)
  • K (number of endmembers) = per scene, heuristic
    The number of materials is chosen by the authors as an approximation of the number of distinct materials in each scene; segmentation quality likely depends on this choice, and no sensitivity analysis is reported (Section 4.6).
  • lambda_spec and lambda_rgb loss weights = lambda_spec=5, lambda_rgb=1
    Loss balancing hyperparameters used in the final objective, Eq. (16).
  • softmax temperature tau = not reported
    The abundance softmax in Eq. (6) uses a temperature tau > 0 but its value is not specified in the paper.
assumptions (4)
  • domain assumption The Extended Linear Mixing Model (ELMM) describes per-point spectral reflectance as a scaled linear combination of a global endmember dictionary.
    Used in Section 4.1 to factor the diffuse component into endmembers, scaling factors, and abundances; assumes scene materials can be captured by a small global dictionary.
  • domain assumption The dichromatic reflection model separates surface reflectance into diffuse and specular components.
    Used in Section 4.2 to combine the unmixed diffuse color with a view-dependent specular term; assumes this two-part decomposition is sufficient for all observed reflectance.
  • ad hoc to paper The cluster probe in Eq. (12) maps a rendered spectral signature to material identity by normalized inner product with the learned endmember dictionary.
    This is a newly introduced segmentation rule, not derived from spectral unmixing theory; it assumes that the dominant endmember in a rendered spectrum indicates the material class.
  • domain assumption Ground-truth material labels for the extended NeSpoF dataset are accurate and meaningful.
    The authors generated these labels from synthetic scenes and use them to evaluate segmentation; the labeling procedure is deferred to the supplementary material.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UnMix-NeRF: Spectral Unmixing Meets Neural Radiance Fields." pith.science (2026). https://pith.science/paper/35REWCQD

@misc{pith2026250621884,
  author       = {Pith},
  title        = {Pith review of: UnMix-NeRF: Spectral Unmixing Meets Neural Radiance Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/35REWCQD}},
  note         = {Machine review of arXiv:2506.21884}
}
read the original abstract

Neural Radiance Field (NeRF)-based segmentation methods focus on object semantics and rely solely on RGB data, lacking intrinsic material properties. This limitation restricts accurate material perception, which is crucial for robotics, augmented reality, simulation, and other applications. We introduce UnMix-NeRF, a framework that integrates spectral unmixing into NeRF, enabling joint hyperspectral novel view synthesis and unsupervised material segmentation. Our method models spectral reflectance via diffuse and specular components, where a learned dictionary of global endmembers represents pure material signatures, and per-point abundances capture their distribution. For material segmentation, we use spectral signature predictions along learned endmembers, allowing unsupervised material clustering. Additionally, UnMix-NeRF enables scene editing by modifying learned endmember dictionaries for flexible material-based appearance manipulation. Extensive experiments validate our approach, demonstrating superior spectral reconstruction and material segmentation to existing methods. Project page: https://www.factral.co/UnMix-NeRF.

Figures

Figures reproduced from arXiv: 2506.21884 by the authors.

Figure 1
Figure 1. UnMix-NeRF: A hyperspectral novel view synthesis framework that leverages spectral unmixing for scene editing and unsuper [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of NeRF unmixing approach. Given a viewing direction (θ, ϕ) and 3D coordinates (x, y, z), our network predicts a density σ and a set of abundance vectors a˜, along with scalar factors ˜s that scale the spectral endmembers E. The scaled endmembers, when multiplied by the predicted abundances a˜, yield the diffuse reflectance component cd. In parallel, a specular reflectance cs and an additional tint factor h… view at source ↗
Figure 3
Figure 3. Visualization of the top-4 performing methods for frame 51 out of 359 for the Caladium plant scene from the Bayspec dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of the learned material abundance maps [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Unsupervised material segmentation of the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 37 canonical work pages

  1. [1]

    Performance comparison of infrared and ul- trasonic sensors for obstacles of different materials in vehi- cle/robot navigation applications

    S Adarsh, S Mohamed Kaleemuddin, Dinesh Bose, and KI Ramachandran. Performance comparison of infrared and ul- trasonic sensors for obstacles of different materials in vehi- cle/robot navigation applications. In IOP Conference Series: Materials Science and Engineering, page 012141. IOP pub- lishing, 2016. 1

    work page 2016

  2. [2]

    Spectral variability in hyperspectral data unmixing: A comprehensive review.IEEE geoscience and remote sens- ing magazine, 9(4):223–270, 2021

    Ricardo Augusto Borsoi, Tales Imbiriba, Jos ´e Carlos Mor- eira Bermudez, C ´edric Richard, Jocelyn Chanussot, Lucas Drumetz, Jean-Yves Tourneret, Alina Zare, and Christian Jutten. Spectral variability in hyperspectral data unmixing: A comprehensive review.IEEE geoscience and remote sens- ing magazine, 9(4):223–270, 2021. 3

    work page 2021

  3. [3]

    Hyperspectral neural radiance fields

    Gerry Chen, Sunil Kumar Narayanan, Thomas Gautier Ot- tou, Benjamin Missaoui, Harsh Muriki, C ´edric Pradalier, and Yongsheng Chen. Hyperspectral neural radiance fields. arXiv preprint arXiv:2403.14839, 2024. 2, 6

    arXiv 2024

  4. [4]

    Torrance

    Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graph- ics (ToG), 1(1):7–24, 1982. 5

    work page 1982

  5. [5]

    Multiple snapshot colored compres- sive spectral imager

    Claudia V Correa, Carlos A Hinojosa, Gonzalo R Arce, and Henry Arguello. Multiple snapshot colored compres- sive spectral imager. Optical Engineering , 56(4):041309– 041309, 2017. 2

    work page 2017

  6. [6]

    Robot-aided hyperspectral imaging for min- eral exploration in underground mining environments

    Sibren Dieters. Robot-aided hyperspectral imaging for min- eral exploration in underground mining environments. 2024. 2

    work page 2024

  7. [7]

    Opennerf: open set 3d neural scene segmentation with pixel- wise features and rendered novel views

    Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Opennerf: open set 3d neural scene segmentation with pixel- wise features and rendered novel views. arXiv preprint arXiv:2404.03650, 2024. 1, 3

    arXiv 2024

  8. [8]

    Sama: Material-aware 3d selection and segmentation

    Michael Fischer, Iliyan Georgiev, Thibault Groueix, Vladimir G Kim, Tobias Ritschel, and Valentin Deschaintre. Sama: Material-aware 3d selection and segmentation. arXiv preprint arXiv:2411.19322, 2024. 3

    arXiv 2024

Show all 48 references
  1. [9]

    Spectral imaging applications: remote sensing, environmental monitoring, medicine, military operations, factory automation, and manufacturing

    Nahum Gat, Suresh Subramanian, Jacob Barhen, and Nikzad Toomarian. Spectral imaging applications: remote sensing, environmental monitoring, medicine, military operations, factory automation, and manufacturing. In 25th AIPR Work- shop: Emerging Applications of Computer Vision, ...

  2. [10]

    Hyperspectral unmix- ing using transformer network

    Preetam Ghosh, Swalpa Kumar Roy, Bikram Koirala, Behnood Rasti, and Paul Scheunders. Hyperspectral unmix- ing using transformer network. IEEE Transactions on Geo- science and Remote Sensing, 60:1–16, 2022. 3

    2022

  3. [11]

    Baking neural ra- diance fields for real-time view synthesis

    Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural ra- diance fields for real-time view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5875–5884, 2021. 4

    2021

  4. [12]

    A fast and accurate similarity-constrained subspace clustering al- gorithm for hyperspectral image

    Carlos Hinojosa, Esteban Vera, and Henry Arguello. A fast and accurate similarity-constrained subspace clustering al- gorithm for hyperspectral image. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 14:10773–10783, 2021. 2

    2021

  5. [13]

    An augmented linear mixing model to ad- dress spectral variability for hyperspectral unmixing

    Danfeng Hong, Naoto Yokoya, Jocelyn Chanussot, and Xiao Xiang Zhu. An augmented linear mixing model to ad- dress spectral variability for hyperspectral unmixing. IEEE Transactions on Image Processing, 28(4):1923–1938, 2018. 3

    1923

  6. [14]

    Nerf-rpn: A general framework for object detection in nerfs

    Benran Hu, Junkai Huang, Yichen Liu, Yu-Wing Tai, and Chi-Keung Tang. Nerf-rpn: A general framework for object detection in nerfs. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 23528–23538, 2023. 1

    2023

  7. [15]

    Spectral imaging with deep learning

    Longqian Huang, Ruichen Luo, Xu Liu, and Xiang Hao. Spectral imaging with deep learning. Light: Science & Ap- plications, 11(1):61, 2022. 2

    2022

  8. [16]

    Lerf: Language embedded radiance fields

    Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 19729–19739,

  9. [17]

    Spectral unmixing

    Nirmal Keshava and John F Mustard. Spectral unmixing. IEEE signal processing magazine, 19(1):44–57, 2002. 2, 3

    2002

  10. [18]

    Neural spectro-polarimetric fields

    Youngchan Kim, Wonjoon Jin, Sunghyun Cho, and Seung- Hwan Baek. Neural spectro-polarimetric fields. In SIG- GRAPH Asia 2023 Conference Papers , pages 1–11, 2023. 2, 6

    2023

  11. [19]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 1, 3

    2023

  12. [20]

    The spectral image processing sys- tem (sips)—interactive visualization and analysis of imaging spectrometer data

    Fred A Kruse, Adam B Lefkoff, Joseph W Boardman, Kathleen B Heidebrecht, AT Shapiro, PJ Barloon, and Alexander FH Goetz. The spectral image processing sys- tem (sips)—interactive visualization and analysis of imaging spectrometer data. Remote sensing of environment, 44(2-3): 1...

    1993

  13. [21]

    Modelling vegetation health and stress using hypersepctral remote sensing data

    Neha Kureel, Jyoti Sarup, Shafique Matin, Suresh Goswami, and Kapil Kureel. Modelling vegetation health and stress using hypersepctral remote sensing data. Modeling Earth Systems and Environment, pages 1–16, 2022. 2

    2022

  14. [22]

    Principles of fluorescence spectroscopy

    Joseph R Lakowicz. Principles of fluorescence spectroscopy. Springer, 2006. 2

    2006

  15. [23]

    Seathru- nerf: Neural radiance fields in scattering media

    Deborah Levy, Amit Peleg, Naama Pearl, Dan Rosenbaum, Derya Akkaynak, Simon Korman, and Tali Treibitz. Seathru- nerf: Neural radiance fields in scattering media. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 56–65, 2023. 2

    2023

  16. [24]

    Spec-nerf: Multi-spectral neural radiance fields

    Jiabao Li, Yuqi Li, Ciliang Sun, Chong Wang, and Jinhui Xiang. Spec-nerf: Multi-spectral neural radiance fields. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 2485–2489. IEEE, 2024. 2

    2024

  17. [25]

    Nerfacc: Efficient sampling accelerates nerfs

    Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. arXiv preprint arXiv:2305.04966, 2023. 5

    2023 arXiv

  18. [26]

    Spectralnerf: Physically based spec- tral rendering with neural radiance field

    Ru Li, Jia Liu, Guanghui Liu, Shengping Zhang, Bing Zeng, and Shuaicheng Liu. Spectralnerf: Physically based spec- tral rendering with neural radiance field. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3154– 3162, 2024. 2

    2024

  19. [27]

    Materialseg3d: Segmenting dense materi- als from 2d priors for 3d assets

    Zeyu Li, Ruitong Gan, Chuanchen Luo, Yuxi Wang, Jiaheng Liu, Ziwei Zhu, Qing Li, Xucheng Yin, Man Zhang, Zhaox- iang Zhang, et al. Materialseg3d: Segmenting dense materi- als from 2d priors for 3d assets. In Proceedings of the 32nd ACM International Conference on Multimedia , ...

    2024

  20. [28]

    Onerf: Unsupervised 3d ob- ject segmentation from multiple views

    Shengnan Liang, Yichen Liu, Shangzhe Wu, Yu-Wing Tai, and Chi-Keung Tang. Onerf: Unsupervised 3d ob- ject segmentation from multiple views. arXiv preprint arXiv:2211.12038, 2022. 3

    2022 arXiv

  21. [29]

    Unsupervised multi-view object segmentation using radiance field propagation

    Xinhang Liu, Jiaben Chen, Huai Yu, Yu-Wing Tai, and Chi- Keung Tang. Unsupervised multi-view object segmentation using radiance field propagation. Advances in Neural Infor- mation Processing Systems, 35:17730–17743, 2022. 3

    2022

  22. [30]

    Sanerf-hq: Segment anything for nerf in high quality

    Yichen Liu, Benran Hu, Chi-Keung Tang, and Yu-Wing Tai. Sanerf-hq: Segment anything for nerf in high quality. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3216–3226, 2024. 1

    2024

  23. [31]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM, 65(1):99–106, 2021. 1, 2, 3

    2021

  24. [32]

    Vertex component analysis: A fast algorithm to unmix hyperspectral data.IEEE transactions on Geoscience and Remote Sensing, 43(4):898– 910, 2005

    Jos ´e MP Nascimento and Jos ´e MB Dias. Vertex component analysis: A fast algorithm to unmix hyperspectral data.IEEE transactions on Geoscience and Remote Sensing, 43(4):898– 910, 2005. 4

    2005

  25. [33]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3

    2023 arXiv

  26. [34]

    Beyond appear- ances: Material segmentation with embedded spectral infor- mation from rgb-d imagery

    Fabian Perez and Hoover Rueda-Chac ´on. Beyond appear- ances: Material segmentation with embedded spectral infor- mation from rgb-d imagery. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 293–301, 2024. 3

    2024

  27. [35]

    Floaters No More: Radiance Field Gradient Scaling for Improved Near-Camera Training

    Julien Philip and Valentin Deschaintre. Floaters No More: Radiance Field Gradient Scaling for Improved Near-Camera Training. In Eurographics Symposium on Rendering . The Eurographics Association, 2023. 5

    2023

  28. [36]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  29. [37]

    Materialistic: Se- lecting similar materials in images

    Prafull Sharma, Julien Philip, Micha ¨el Gharbi, Bill Freeman, Fredo Durand, and Valentin Deschaintre. Materialistic: Se- lecting similar materials in images. ACM Transactions on Graphics, 42(4), 2023. 3

    2023

  30. [38]

    Nerfstudio: A modular framework for neural radiance field development

    Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 conference proceedings, pages 1– 12...

    2023

  31. [39]

    Lidar-nerf: Novel li- dar view synthesis via neural radiance fields

    Tang Tao, Longfei Gao, Guangrun Wang, Yixing Lao, Peng Chen, Hengshuang Zhao, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. Lidar-nerf: Novel li- dar view synthesis via neural radiance fields. In Proceedings of the 32nd ACM International Conference on Multimedi...

    2024

  32. [40]

    Hypergs: Hyperspectral 3d gaussian splatting

    Christopher Thirgood, Oscar Mendez, Erin Chao Ling, Jon Storey, and Simon Hadfield. Hypergs: Hyperspectral 3d gaussian splatting. arXiv preprint arXiv:2412.12849, 2024. 2

    2024 arXiv

  33. [41]

    Matfuse: controllable material gen- eration with diffusion models

    Giuseppe Vecchio, Renato Sortino, Simone Palazzo, and Concetto Spampinato. Matfuse: controllable material gen- eration with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4429–4438, 2024. 1

    2024

  34. [42]

    A new extended linear mixing model to address spectral variability

    Miguel Angel Veganzones, Lucas Drumetz, Guillaume To- chon, Mauro Dalla Mura, Antonio Plaza, J Bioucas-Dias, and Jocelyn Chanussot. A new extended linear mixing model to address spectral variability. In 2014 6th Workshop on Hy- perspectral Image and Signal Processing: Evolutio...

    2014

  35. [43]

    Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields

    Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 5481–5490. IE...

    2022

  36. [44]

    Mapping spectrally similar urban materials at sub-pixel scales

    Erin B Wetherley, Dar A Roberts, and Joseph P McFad- den. Mapping spectrally similar urban materials at sub-pixel scales. Remote Sensing of Environment, 195:170–183, 2017. 2

    2017

  37. [45]

    Thermal-nerf: Neural radiance fields from an infrared cam- era

    Tianxiang Ye, Qi Wu, Junyuan Deng, Guoqing Liu, Liu Liu, Songpengcheng Xia, Liang Pang, Wenxian Yu, and Ling Pei. Thermal-nerf: Neural radiance fields from an infrared cam- era. arXiv preprint arXiv:2403.10340, 2024. 2

    2024 arXiv

  38. [46]

    Hyperspectral unmixing via deep convolu- tional neural networks

    Xiangrong Zhang, Yujia Sun, Jingyan Zhang, Peng Wu, and Licheng Jiao. Hyperspectral unmixing via deep convolu- tional neural networks. IEEE Geoscience and Remote Sens- ing Letters, 15(11):1755–1759, 2018. 3

    2018

  39. [47]

    A laboratory-created dataset with ground truth for hyperspectral unmixing eval- uation

    Min Zhao, Jie Chen, and Zhe He. A laboratory-created dataset with ground truth for hyperspectral unmixing eval- uation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 12(7):2170–2183, 2019. 2

    2019

  40. [48]

    Materobot: Material recognition in wearable robotics for people with visual impairments

    Junwei Zheng, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Materobot: Material recognition in wearable robotics for people with visual impairments. In 2024 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 2303–2309. IEEE, 2024. 1

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.