Pith. sign in

REVIEW 2 major objections 6 minor 50 references

This paper shows that polarized light transport can be differentiated without inverting rank-deficient Mueller matrices by caching suffix radiance and replaying paths, yielding unbiased, memory-efficient gradients for inverse rendering.

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

2026-08-02 05:39 UTC pith:METPLGS3

load-bearing objection A genuinely useful extension of path replay to polarized transport with a sound cached-suffix core, but the hybrid variant's threshold-dependent branching is unproven and underspecified; worth refereeing. the 2 major comments →

arxiv 2607.13265 v1 pith:METPLGS3 submitted 2026-07-14 cs.CV

Differentiable Polarized Path Tracing

Pramod Rao , J\'er\'emy Riviere , Xilong Zhou , Abhijeet Ghosh , Abhimitra Meka , Thabo Beeler , Marc Habermann , Christian Theobalt
show 1 more author
Delio Vicini
This is my paper
classification cs.CV
keywords differentiable renderingpolarized light transportpath replay backpropagationMueller-Stokes calculusinverse renderinggradient estimationStokes vectors
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.

Inverse rendering needs gradients of a rendered image with respect to scene parameters. For unpolarized light, path replay backpropagation computes these gradients in constant memory, but it relies on dividing by the local BSDF value at each bounce—and in polarized rendering that value is a 4x4 Mueller matrix that is often singular. The paper's central claim is that this division can be replaced by a lookup: store the accumulated outgoing radiance at each path vertex during a forward pass, then replay the same random path and read the stored suffix directly. This removes the need to invert rank-deficient Mueller operators, produces gradients that match automatic differentiation within noise, and keeps memory roughly constant up to depth 32. A sympathetic reader would care because it makes polarization—which carries extra geometric and material constraints—practical for gradient-based optimization.

Core claim

In polarized rendering, each path vertex evaluates a BSDF as a 4x4 Mueller matrix M; the standard path-replay adjoint pass removes the contribution of a sampled interaction by multiplying by M^{-1}. But common polarimetric elements—diffuse reflections, linear polarizers—map Stokes space to a lower-dimensional subspace, so M is singular. The paper discovers that the inversion is avoidable: if the primal pass stores the suffix radiance L[i+1] (the light arriving after the vertex) computed by a backward fold of detached throughputs, the adjoint pass can replay the same random path and multiply the differentiated local Mueller matrix by the cached suffix. The authors argue this is unbiased, veri

What carries the argument

Cached suffix replay: a two-pass algorithm. The primal pass records per-vertex emission and throughput (as 4x4 Mueller matrices) and reconstructs suffix radiance by a backward fold; the adjoint pass replays paths from the same random seeds and at each interaction loads the cached suffix radiance L[i+1] instead of computing M^{-1} L. A hybrid extension checkpoints every k vertices and decides suffix recovery using the depolarization index (a scalar measuring loss of polarization) and a determinant threshold, falling back to scalar PRB for effectively depolarizing blocks or to recursive recomputation for near-singular ones. This mechanism carries the argument because it eliminates the only pla

Load-bearing premise

The cached suffix radiance computed in the primal pass is treated as a detached constant in the adjoint replay, and the paper asserts—without a proof for the polarized case—that this introduces no bias; if that factorization fails for partially depolarizing or near-singular Mueller operators, the gradients would be biased.

What would settle it

Compare this method's gradients against a finite-difference reference of the forward polarized renderer in a scene whose Mueller matrices are nearly singular but not exactly (e.g., a polarizer with small diattenuation error, or a rough dielectric). If the gradient bias grows as the smallest singular value of the sampled Mueller matrix approaches zero—or if the cached-suffix estimator disagrees with the finite-difference derivative beyond Monte Carlo noise—the unbiasedness claim is falsified. A simpler version: nest several polarizers so the suffix radiance depends strongly on a differentiated

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Gradients with respect to a linear polarizer's rotation angle become usable, enabling applications like glare reduction by optimizing an optical element's orientation.
  • Polarization-aware optimization recovers diffuse texture, roughness, and normal maps with lower error than intensity-only baselines, because Stokes degrees of freedom carry complementary constraints.
  • Memory usage stays approximately constant in path depth up to depth 32, scaling to deeper paths via the hybrid cache–recompute variant, so full polarized differentiation generalizes to complex scenes.
  • The method combines with visibility-aware projective sampling to improve single-view 3D reconstruction, lowering Chamfer distance and normal error relative to unpolarized geometry optimization.

Where Pith is reading between the lines

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

  • A natural extension is to other singular transport operators—participating media with degenerate phase matrices, or BSSRDFs with depolarizing subsurface scattering—where the same cached-suffix idea could avoid ill-defined inverses.
  • The paper validates unbiasedness empirically; a formal proof would need to show that the cached suffix, though computed without differentiating through it, is a valid detached estimate under correlated path sampling. A careful stress test with partially depolarizing, near-singular Mueller matrices would settle it.
  • The hybrid variant's threshold heuristics (depolarization index below a threshold falls back to scalar PRB) introduce a non-smooth switch; if the threshold is crossed during optimization, gradient discontinuities could appear—worth testing in practice.
  • Since Stokes/Mueller tensors are 4x4 per wavelength channel, the memory/compute tradeoff differs between RGB and spectral polarized rendering; the method's advantage would presumably grow with channel count, an implication the paper doesn't draw.

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

2 major / 6 minor

Summary. This paper presents a differentiable polarized path tracing method. The authors extend path replay backpropagation (PRB) to Mueller-Stokes calculus, showing that rank-deficient Mueller matrices break the inversion step of PRB. They propose 'cached suffix replay' (Algorithm 3), which stores per-vertex suffix radiance to avoid inversion, and a 'hybrid cached replay' (Section 4.4) that combines block-wise caching with recursive recomputation. The method is validated by comparing reverse-mode gradients against conventional automatic differentiation (AD) in Mitsuba 3, and demonstrated on inverse rendering tasks (polarizer angle optimization, texture/normal recovery) and geometry reconstruction, with claims of lower memory and runtime than baselines.

Significance. If the central claims hold, the fully cached algorithm is a practical contribution: it provides a memory-efficient way to differentiate polarized light transport, enabling polarization-aware inverse rendering tasks that are infeasible with conventional AD. The paper includes careful comparisons with conventional AD, and the core cached replay idea is elegant and appears algebraically sound. However, the unbiasedness claim is asserted for the overall approach, while the hybrid variant is not proven or validated, and the manuscript omits the promised supplemental details. The free parameters γ and ε also lack guidance. These gaps currently prevent acceptance.

major comments (2)
  1. [§4.4, Eq. (3)] The hybrid replay rule switches among scalar PRB, matrix inversion, and recursive recomputation based on DI(β̄)<γ and |det(β̄)|>ε. Since β̄ is a function of the differentiable scene parameters, the branch decision is parameter-dependent. Unbiasedness of each branch does not imply unbiasedness of the piecewise estimator: differentiating the interior of each branch omits boundary contributions at the threshold surfaces, exactly the visibility-boundary problem in differentiable rendering. No proof, convergence analysis, or empirical validation of the hybrid's gradient unbiasedness is provided; full pseudocode and benchmarks are deferred to a supplemental that is absent from this manuscript. Thus the paper's claim of 'unbiased gradients' (abstract, §4.3) is not established for the memory-reduced variant, and the method is incompletely specified.
  2. [§5.4, Fig. 7] The paper states 'We also evaluate the hybrid variant introduced in Sec. 4.4' (Sec. 5.4) and that it 'becomes useful at larger depths,' but the manuscript presents no results for the hybrid: Fig. 7 shows only memory/time for the fully cached method and baselines, not gradient correctness or bias for the hybrid. The hybrid's practical accuracy is therefore unsubstantiated, and its thresholds γ and ε are not analyzed or tuned. This is a load-bearing gap because the hybrid is part of the claimed contribution and the unbiasedness claim is stated broadly.
minor comments (6)
  1. [§5, Fig. 7 caption] The caption mentions 'NVIDIA GeForce RTX 5090' while the text in §5 says all evaluations were run on an RTX 4090. Please reconcile this discrepancy.
  2. [§4.4, Eq. (3)] The depolarization index DI(·) is not defined. Please provide its formula or a reference.
  3. [§4.4] The notation β̄ (block throughput) is introduced but not formally defined. The relationship between β̄ and the per-vertex throughput β in Algorithm 3 should be clarified.
  4. [§5.1, Fig. 3] The 'RE' metric is described as 'relative error' but the formula is not given. Please specify how it is computed (e.g., L2 error normalized by reference norm).
  5. [Algorithm 3] Line 2 initializes L and β as lists of zero/identity matrices; the dimensions (4×4) and the fact that L[j] is a list of lists could be stated more explicitly for readability.
  6. [§4.2] The noise regularization from [37] is said to add u·I_4 to M. The distribution of u and the handling of the same noise in the adjoint phase are not detailed; a brief description would help reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the cached suffix replay estimator is defined by the path contribution itself and is checked against an independent AD reference; deferred hybrid pseudocode/proofs are completeness gaps, not circular reductions.

full rationale

The derivation chain is self-contained. Algorithm 3 defines the suffix cache by a backward fold L[j] += beta[j]*L[j+1] and the adjoint pass accumulates backward(deltaL * beta * M * L[i+1]); this is the product-rule expansion of the path estimator, not a fitted or renamed input. The paper's central claim of unbiased gradients is validated in Sec. 5.1 against Mitsuba 3 conventional AD, an external reference, and no threshold or model parameter is reported as fitted to those gradient comparisons. The most exposed overlapping-author citation is [37] (Path Replay Backpropagation), which is used as the foundation to extend; it is not the target result, and the extension's correctness is checked externally. The hybrid variant in Sec. 4.4 (Eq. 3) introduces threshold-dependent branches and the text states 'Full pseudocode and extended benchmarks are provided in the supplemental material' (Sec. 4.4) and 'Detailed plots are included in the supplemental material' (Sec. 5.4); these are omitted-proof/completeness limitations that could hide a boundary-bias issue, but they are not a circular equivalence. No self-definitional definitions, fitted-input predictions, imported uniqueness theorems, or ansatz-by-citation were found.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The fully cached replay (Algorithm 3) has no fitted parameters; the hybrid variant introduces hand-set thresholds γ and ε (Eq. 3). The core assumption is that cached suffix radiance can be detached during adjoint replay without bias. No new physical entities are introduced.

free parameters (2)
  • γ (depolarization-index threshold)
    Appears in Eq. (3) controlling when the scalar PRB update is used in hybrid replay; chosen by hand, no sensitivity analysis provided. Only affects the hybrid variant, not the fully cached method.
  • ε (determinant admissibility threshold)
    Appears in Eq. (3) controlling when local matrix inversion is attempted; chosen by hand, no sensitivity analysis provided. Affects only the hybrid variant.
axioms (4)
  • standard math Path replay backpropagation produces unbiased gradients in the unpolarized setting and can be extended by local inversion where Mueller matrices are invertible.
    Borrowed from Vicini et al. [37]; the paper relies on this as the foundation for extending to polarized transport (§3.2).
  • domain assumption Mueller–Stokes calculus and Mitsuba 3 coordinate-frame conversions are correct and all incident/outgoing Stokes vectors are expressed in consistent frames.
    Polarized rendering requires consistent coordinate frames; paper assumes Mitsuba implementation handles conversions (§3.3).
  • domain assumption Ideal depolarizers need no incident polarization information; diffuse Mueller matrices of rank 1 can be handled by scalar PRB.
    §4.1 Case 2; this is physically motivated but not proven in the paper.
  • ad hoc to paper Cached suffix radiance from the primal pass can be treated as a detached constant in the adjoint pass without introducing bias.
    Central algorithmic trick (§4.3, Algorithm 3); no formal proof given, only empirical validation against AD.

pith-pipeline@v1.3.0-alltime-deepseek · 12388 in / 10793 out tokens · 109663 ms · 2026-08-02T05:39:57.332232+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Differentiable Polarized Path Tracing." pith.science (2026). https://pith.science/paper/METPLGS3

@misc{pith2026260713265,
  author       = {Pith},
  title        = {Pith review of: Differentiable Polarized Path Tracing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/METPLGS3}},
  note         = {Machine review of arXiv:2607.13265}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Physically based differentiable rendering has proven to be a powerful tool for inverse rendering problems (e.g., 3D reconstruction, reflectance estimation, lighting estimation). However, most existing methods operate solely on radiometric intensity, discarding valuable polarization cues that constrain scene geometry and material properties. While forward simulation of polarized light is well-defined via Mueller-Stokes calculus, extending reverse-mode differentiation to this domain presents significant challenges. The rank-deficient nature of common polarimetric operators, such as linear polarizers and diffuse reflections, violates the invertibility assumptions of standard gradient estimators like path replay backpropagation and results in numerical instability. We address this by proposing a robust, polarization-aware differentiable path tracing method. Our approach estimates unbiased gradients through a combination of path replay and local caching. This formulation enables efficient and stable optimization of material and lighting parameters in complex scenes, broadening the applicability of physically based inverse rendering. Project page: https://vcai.mpi-inf.mpg.de/projects/DPPT/

Figures

Figures reproduced from arXiv: 2607.13265 by Abhijeet Ghosh, Abhimitra Meka, Christian Theobalt, Delio Vicini, J\'er\'emy Riviere, Marc Habermann, Pramod Rao, Thabo Beeler, Xilong Zhou.

Figure 1
Figure 1. Figure 1: Polarization plays an important role in differentiable inverse rendering. (a) Pa￾rameters that control polarization effects such as a linear polarizer rotation θ can pro￾duce significant gradient contributions, but existing differentiable path tracing methods struggle to compute these gradients efficiently and accurately. We introduce a caching mechanism for polarized path replay that avoids inverting rank… view at source ↗
Figure 2
Figure 2. Figure 2: Example for Cached Suffix Replay on a light path of depth 2. The primal pass accumulates the full path contribution, tracks the detached BSDF contributions βi, and constructs the cache of suffix contributions via a backward fold. The adjoint pass then directly re-loads the incident polarization state and multiplies it by the differentiated BSDF weight. This enables unbiased gradient estimation without any … view at source ↗
Figure 3
Figure 3. Figure 3: Gradient Analysis. We visualize gradient images and compare against the Conv. AD reference. Columns report P-PRB, P-RB, and our method, with each gra￾dient tile annotated by its relative error (RE) w.r.t. Conv. AD (preferred over RMSE to better account for small gradient magnitudes). Rows correspond to different scenes with their associated parameters listed on the right. We use conventional AD as the refe… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of reverse-mode gradients under different polarization configurations. We render a virtual light stage setup with right-circular illumination and accumulate gradients over a multi-view camera setup. Columns correspond to polarizer configurations (0 ◦ , 45◦ , 90◦ , and left-circular), and rows show gradients with respect to different scene parameters (surface normals, roughness, and s… view at source ↗
Figure 5
Figure 5. Figure 5: Polarized inverse rendering. We evaluate four inverse rendering tasks (rows) and compare conventional automatic differentiation (Conv. AD), unpolarized PRB [37], and our polarization-aware method against the ground truth reference. From left to right, we show the initial state, the optimized result using Conv. AD, PRB, our method, and the reference. The tasks include optimizing the rotation angle of a line… view at source ↗
Figure 6
Figure 6. Figure 6: Single-view 3D reconstruction. Ours (Pol.) reconstructed geometry is sharper than the unpolarized baseline (Unpol.), shown from test views. 5.3 Geometry Reconstruction with Polarization Classical shape-from-polarization methods exploit the fact that the polarization state of reflected light carries surface-orientation information. To demonstrate this, we further extend the projective sampling integrator [4… view at source ↗
Figure 7
Figure 7. Figure 7: Performance evaluation on a polarized scene rendered at 1280 × 720 resolution with 1 sample per pixel. (a) We use the Modern Hall scene and introduce a stack of 85 linear polarizers to simulate long light paths. (b) We illustrate the effect of increasing the maximum path depth on the rendered appearance. We then compute gradients with respect to the floor albedo using conventional automatic differentiation… 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

50 extracted references · 3 canonical work pages

  1. [1]

    In: IEEE International Conference on Computer Vision (ICCV)

    Atkinson, G.A., Hancock, E.R.: Multi-view surface reconstruction using polariza- tion. In: IEEE International Conference on Computer Vision (ICCV). vol. 1, pp. 309–316. IEEE (2005)

    2005

  2. [2]

    IEEE transactions on image processing15(6), 1653–1664 (2006)

    Atkinson, G.A., Hancock, E.R.: Recovery of surface orientation from diffuse polar- ization. IEEE transactions on image processing15(6), 1653–1664 (2006)

    2006

  3. [3]

    Ba, Y., Gilbert, A., Wang, F., Yang, J., Chen, R., Wang, Y., Yan, L., Shi, B., Kadambi,A.:Deepshapefrompolarization.In:EuropeanConferenceonComputer Vision (ECCV). pp. 554–571. Springer (2020)

    2020

  4. [4]

    ACM Trans

    Baek, S.H., Jeon, D.S., Tong, X., Kim, M.H.: Simultaneous acquisition of polari- metric svbrdf and normals. ACM Trans. Graph. (Proc. SIGGRAPH Asia)36(6), 268:1–15 (2018).https://doi.org/10.1145/3272127.3275018,http://dx.doi.org/ 10.1145/3272127.3275018

    arXiv 2018

  5. [5]

    ACM Trans

    Baek, S.H., Zeltner, T., Ku, H.J., Hwang, I., Tong, X., Jakob, W., Kim, M.H.: Image-based acquisition and modeling of polarimetric reflectance. ACM Trans. Graph. (Proc. SIGGRAPH)39(4) (2020)

    2020

  6. [6]

    Bangaru, S., Li, T.M., Durand, F.: Unbiased warped-area sampling for differen- tiablerendering.ACMTrans.Graph.(Proc.SIGGRAPHAsia)39(6),245:1–245:18 (2020).https://doi.org/10.1145/3414685.3417833

    arXiv 2020

  7. [7]

    In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Boult, T.E., Wolff, L.B.: Physically-based edge labeling. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 656–662 (1991)

    1991

  8. [8]

    SPIE Publications (2005)

    Collet, E.: Field Guide to Polarization. SPIE Publications (2005)

    2005

  9. [9]

    In: IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR) (June 2021)

    Deschaintre, V., Lin, Y., Ghosh, A.: Deep polarization imaging for 3d shape and svbrdf acquisition. In: IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR) (June 2021)

    2021

  10. [10]

    ACM Transactions on Graphics (TOG)29(6), 1–12 (2010)

    Ghosh, A., Chen, T., Peers, P., Wilson, C.A., Debevec, P.: Circularly polarized spherical illumination reflectometry. ACM Transactions on Graphics (TOG)29(6), 1–12 (2010)

    2010

  11. [11]

    ACM Trans

    Ghosh, A., Fyffe, G., Tunwattanapong, B., Busch, J., Yu, X., Debevec, P.: Mul- tiview face capture using polarized spherical gradient illumination. ACM Trans. Graph. (Proc. SIGGRAPH Asia)30(6), 1–10 (Dec 2011).https://doi.org/10. 1145/2070781.2024163

    arXiv 2011

  12. [12]

    In: European Conference on Com- puter Vision (ECCV)

    Gkioulekas, I., Levin, A., Zickler, T.: An evaluation of computational imaging techniques for heterogeneous inverse scattering. In: European Conference on Com- puter Vision (ECCV). pp. 685–701. Springer International Publishing (2016). https://doi.org/10.1007/978-3-319-46487-9_42

    work page doi:10.1007/978-3-319-46487-9_42 2016

  13. [13]

    Griewank, A., Walther, A.: Evaluating derivatives: principles and techniques of algorithmic differentiation, vol. 105. SIAM (2008)

    2008

  14. [14]

    arXiv preprint arXiv:2512.00794 (2025)

    Guo, B., Wen, S., Zhao, Y., Li, J., Zheng, Z.: Polargs: Polarimetric cues for ambiguity-free gaussian splatting with accurate geometry recovery. arXiv preprint arXiv:2512.00794 (2025)

    arXiv 2025

  15. [15]

    SIGGRAPH Asia 2024)43(6) (2024)

    Ha, H., Hwang, I., Monzon, N., Cho, J., Kim, D., Baek, S.H., Muñoz, A., Gutierrez, D.,Kim,M.H.:Polarimetricbssrdfacquisitionofdynamicfaces.ACMTransactions on Graphics (Proc. SIGGRAPH Asia 2024)43(6) (2024)

    2024

  16. [16]

    In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Han, Y., Guo, H., Fukai, K., Santo, H., Shi, B., Okura, F., Ma, Z., Jia, Y.: Nersp: Neural 3d reconstruction for reflective objects with sparse polarized im- ages. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11821–11830 (2024) 18 P. Rao et al

    2024

  17. [17]

    In: IEEE International Conference on Computer Vision (ICCV)

    Han, Y., Tie, B., Guo, H., Lyu, Y., Li, S., Shi, B., Jia, Y., Ma, Z.: Polgs: Po- larimetric gaussian splatting for fast reflective surface reconstruction. In: IEEE International Conference on Computer Vision (ICCV). pp. 28073–28082 (2025)

    2025

  18. [18]

    ACM Trans

    Hwang, I., Jeon, D.S., Muñoz, A., Gutierrez, D., Tong, X., Kim, M.H.: Sparse ellip- sometry: Portable acquisition of polarimetric svbrdf and shape with unstructured flash photography. ACM Trans. Graph. (Proc. SIGGRAPH)41(4) (2022)

    2022

  19. [19]

    Jakob, W., Speierer, S., Roussel, N., Nimier-David, M., Vicini, D., Zeltner, T., Nicolet, B., Crespo, M., Leroy, V., Zhang, Z.: Mitsuba 3 renderer (2022), https://mitsuba-renderer.org

    2022

  20. [20]

    In: IEEE International Conference on Computer Vision (ICCV)

    Kadambi, A., Taamazyan, V., Shi, B., Raskar, R.: Polarized 3d: High-quality depth sensing with polarization cues. In: IEEE International Conference on Computer Vision (ICCV). pp. 3370–3378 (2015)

    2015

  21. [21]

    Computer Graphics (Proc

    Kajiya, J.T.: The rendering equation. Computer Graphics (Proc. SIGGRAPH) 20(4), 143–150 (Aug 1986).https://doi.org/10/cvf53j

    1986

  22. [22]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. (Proc. SIGGRAPH)42(4), 139–1 (2023)

    2023

  23. [23]

    In: Proceedings of the 6th international joint conference on Artificial intelligence- Volume 1

    Koshikawa, K.: A polarimetric approach to shape understanding of glossy objects. In: Proceedings of the 6th international joint conference on Artificial intelligence- Volume 1. pp. 493–495 (1979)

    1979

  24. [24]

    In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Lei, C., Qi, C., Xie, J., Fan, N., Koltun, V., Chen, Q.: Shape from polarization for complex scenes in the wild. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12632–12641 (2022)

    2022

  25. [25]

    ACM Trans

    Li, T.M., Aittala, M., Durand, F., Lehtinen, J.: Differentiable monte carlo ray tracing through edge sampling. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 37(6), 222:1–222:11 (2018).https://doi.org/10.1145/3272127.3275109

    arXiv 2018

  26. [26]

    In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers

    Lin, A., Ghosh, A.: Single-shot facial capture using polarized rgb sinusoidal illu- mination. In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers. SA Conference Papers ’25, Association for Computing Machinery, New York, NY, USA (2025).https://doi.org/10.1145/3757377.3763942,https://doi.org/10. 1145/3757377.3763942

    arXiv 2025

  27. [27]

    In: Kautz, J., Pattanaik, S

    Ma, W.C., Hawkins, T., Peers, P., Chabert, C.F., Weiss, M., Debevec, P.: Rapid Acquisition of Specular and Diffuse Normal Maps from Polarized Spher- ical Gradient Illumination. In: Kautz, J., Pattanaik, S. (eds.) Rendering Tech- niques (Proc. EG Symposium on Rendering). Eurographics Association (2007). https://doi.org/10.2312/EGWR/EGSR07/183-194

    work page doi:10.2312/egwr/egsr07/183-194 2007

  28. [28]

    In: Euro- pean Conference on Computer Vision (ECCV) (2020).https://doi.org/10.1145/ 3503250

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. In: Euro- pean Conference on Computer Vision (ECCV) (2020).https://doi.org/10.1145/ 3503250

    2020

  29. [29]

    In: IEEE International Conference on Computer Vision (ICCV)

    Miyazaki, Kagesawa, Ikeuchi: Polarization-based transparent surface modeling from two views. In: IEEE International Conference on Computer Vision (ICCV). pp. 1381–1386. IEEE (2003)

    2003

  30. [30]

    Miyazaki, D., Shigetomi, T., Baba, M., Furukawa, R., Hiura, S., Asada, N.: Polarization-based surface normal estimation of black specular objects from multi- pleviewpoints.In:2012SecondInternationalConferenceon3DImaging,Modeling, Processing, Visualization & Transmission. pp. 104–111. IEEE (2012)

    2012

  31. [31]

    ACM Trans

    Nimier-David, M., Speierer, S., Ruiz, B., Jakob, W.: Radiative backpropagation: An adjoint method for lightning-fast differentiable rendering. ACM Trans. Graph. (Proc. SIGGRAPH)39(4) (Jul 2020).https://doi.org/10.1145/3386569.3392406 Differentiable Polarized Path Tracing 19

    arXiv 2020

  32. [32]

    ACM Trans

    Nimier-David, M., Vicini, D., Zeltner, T., Jakob, W.: Mitsuba 2: A retargetable forward and inverse renderer. ACM Trans. Graph. (Proc. SIGGRAPH Asia)38(6) (Nov 2019).https://doi.org/10.1145/3355089.3356498

    arXiv 2019

  33. [33]

    In: Ghosh, A., Wei, L.Y

    Nogue, E., Lin, Y., Ghosh, A.: Polarization-imaging Surface Reflectometry using Near-field Display. In: Ghosh, A., Wei, L.Y. (eds.) Comp. Graph. Forum (Proc. EGSR). The Eurographics Association (2022).https://doi.org/10.2312/sr. 20221154

    work page doi:10.2312/sr 2022

  34. [34]

    In: ACM SIG- GRAPH ASIA 2025 Conference Proceedings (2025)

    Rao, P., Meka, A., Zhou, X., Fox, G., B R, M., Zhan, F., Weyrich, T., Bickel, B., Pfister, H., Matusik, W., Beeler, T., Elgharib, M., Habermann, M., Theobalt, C.: 3dpr: Single image 3d portrait relighting with generative priors. In: ACM SIG- GRAPH ASIA 2025 Conference Proceedings (2025)

    2025

  35. [35]

    ACM Trans

    Riviere, J., Reshetouski, I., Filipi, L., Ghosh, A.: Polarization imaging reflectome- try in the wild. ACM Trans. Graph. (Proc. SIGGRAPH Asia)36(6) (Nov 2017). https://doi.org/10.1145/3130800.3130894,https://doi.org/10.1145/3130800. 3130894

    arXiv 2017

  36. [36]

    CoRR abs/2006.15059(2020),https://arxiv.org/abs/2006.15059

    Stam, J.: Computing light transport gradients using the adjoint method. CoRR abs/2006.15059(2020),https://arxiv.org/abs/2006.15059

    Pith/arXiv arXiv 2006

  37. [37]

    ACM Trans

    Vicini, D., Speierer, S., Jakob, W.: Path replay backpropagation: Differentiating light paths using constant memory and linear time. ACM Trans. Graph. (Proc. SIGGRAPH)40(4) (Aug 2021).https://doi.org/10.1145/3450626.3459804

    arXiv 2021

  38. [38]

    In: Comp

    Wanaset, R., Guarnera, C., Smith, W.A.P.: Neural field multi-view shape-from- polarisation. In: Comp. Graph. Forum. vol. 44. Wiley-Blackwell (2025)

    2025

  39. [39]

    ACM Trans

    Weier, P., Riviere, J., Guseinov, R., Garbin, S., Slusallek, P., Bickel, B., Beeler, T., Vicini, D.: Practical inverse rendering of textured and translucent appearance. ACM Trans. Graph. (Proc. SIGGRAPH)44(4) (Aug 2025).https://doi.org/10. 1145/3730855

    2025

  40. [40]

    In: Proceedings of a workshop on Image un- derstanding

    Wolff, L.B.: Material classification and separation of reflection components using polarization/radiometric information. In: Proceedings of a workshop on Image un- derstanding. pp. 232–244 (1989)

    1989

  41. [41]

    In: IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR)

    Wolff, L.B.: Using polarization to separate reflection components. In: IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 363–369. IEEE (1989)

    1989

  42. [42]

    IEEE Transactions on Pattern Analysis and Machine Intelligence12(11), 1059– 1071 (1990)

    Wolff, L.B.: Polarization-based material classification from specular reflection. IEEE Transactions on Pattern Analysis and Machine Intelligence12(11), 1059– 1071 (1990)

    1990

  43. [43]

    In: Optics, Illumination, and Image Sensing for Machine Vision IV

    Wolff, L.B.: Surface orientation from two camera stereo with polarizers. In: Optics, Illumination, and Image Sensing for Machine Vision IV. vol. 1194, pp. 287–297. SPIE (1990)

    1990

  44. [44]

    In: SIGGRAPH Asia Conference Papers (2024)

    Xu, P., Bangaru, S., Li, T.M., Zhao, S.: Markov-chain monte carlo sampling of visibility boundaries for differentiable rendering. In: SIGGRAPH Asia Conference Papers (2024)

    2024

  45. [45]

    Yang, X., Liu, Z., Nie, Z., Fan, Q., Shi, Z., Bonar, J., Heidrich, W.: End-to-end differentiable design of geometric waveguide displays (2026)

    2026

  46. [46]

    arXiv preprint arXiv:2601.04370 (2026)

    Yang, X., Liu, Z., Nie, Z., Fan, Q., Shi, Z., Bonar, J., Heidrich, W.: End- to-end differentiable design of geometric waveguide displays. arXiv preprint arXiv:2601.04370 (2026)

    arXiv 2026

  47. [47]

    readthedocs.io/en/stable/src/key_topics/polarization.html(2021), [Online; accessed 16-January-2026] 20 P

    Zeltner, T.: Mitsuba 3 documentation: Polarization.https : / / mitsuba . readthedocs.io/en/stable/src/key_topics/polarization.html(2021), [Online; accessed 16-January-2026] 20 P. Rao et al

    2021

  48. [48]

    ACM Trans

    Zhang, C., Wu, L., Zheng, C., Gkioulekas, I., Ramamoorthi, R., Zhao, S.: A differ- ential theory of radiative transfer. ACM Trans. Graph. (Proc. SIGGRAPH Asia) 38(6) (Nov 2019).https://doi.org/10.1145/3355089.3356522

    arXiv 2019

  49. [49]

    ACM Trans

    Zhang, Z., Roussel, N., Jakob, W.: Projective sampling for differentiable rendering of geometry. ACM Trans. Graph. (Proc. SIGGRAPH Asia)42(6) (Dec 2023). https://doi.org/10.1145/3618385

    doi:10.1145/3618385 2023

  50. [50]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Zhou, X., Chen, J., Rao, P., Teufel, T., Lyu, L., Minasian, T., Sotnychenko, O., Long, X.X., Habermann, M., Theobalt, C.: Olatverse: A large-scale real-world ob- ject dataset with precise lighting control. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 28848–28859 (2026)

    2026