Pith. sign in

REVIEW 3 major objections 6 minor 54 references

Deep Unrolling of Sparsity-Induced RDO for 3D Point Cloud Attribute Coding

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that unrolling conjugate gradient descent and proximal gradient descent into feed-forward networks yields a practical, trainable, generalized RAHT transform whose rate-distortion-optimized encoder reduces bit rate by 6–11%

desk verdict Useful unrolled RAHT paper whose headline 6–11% gain rests on an estimated G-PCC baseline; send to review with requests for validation and code. read the letter →

arxiv 2509.08685 v2 pith:AOWXPWI7 submitted 2025-09-10 eess.IV cs.ITcs.LGmath.IT

Tam Thuc Do , Philip A. Chou , Gene Cheung This is my paper · ORCID
classification eess.IVcs.ITcs.LGmath.IT
keywords pointcloudattributecodingalgorithmunrollingrate-distortionoptimizationB-splinesRAHTproximalgradientdescentconjugatelearnedcompression
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the expensive, geometry-dependent matrix inversions in higher-order RAHT can be replaced by unrolled iterative solvers, making the transform linear in the number of points and trainable end-to-end. The encoder is built by unrolling an accelerated proximal gradient descent algorithm that minimizes a squared-error distortion plus an L1 sparsity penalty, a form that follows from modeling the transform coefficients as i.i.d. zero-mean Laplacians. If correct, a decoder-only compression scheme with very few learned parameters can outperform the MPEG G-PCC predictor, achieving 6–11% bit rate reduction at equal PSNR on Longdress, Redandblack, Loot, and Soldier. The paper further claims this is the first generalized RAHT that is both practical (linear complexity) and trainable.

What carries the argument

The accelerating PGD encoder with the shrinkage (soft-thresholding) operator, whose unrolled iterations become trainable layers, is the engine of the encoder. The decoder is made linear-complexity by unrolling conjugate gradient descent for (ΦᵀΦ)⁻¹ and a Taylor series for (ΦᵀΦ)⁻¹/², both operating on the sparse Gram matrix. The prediction module M, built from learned operators B_l, recasts the coefficients as prediction residuals that are easier to entropy code.

What would settle it

Measure the empirical histograms of the unquantized transform coefficients produced by the learned encoder at each resolution level and compare them to the fitted Laplacian distributions used in the rate model; a strong mismatch (heavy tails, bimodality, or level-dependent shapes that the per-level scale cannot capture) would refute the claim that the encoder minimizes the true rate-distortion objective, and would likely show up as a growing gap between the training proxy rate and the RLGR-coded bit rate.

Watch

Extended reading notes

Core claim

For a given decoder basis Θ, the paper shows that the rate-distortion optimal encoder reduces to minimizing ||f − ΘV||² + λ||ΓV||₁ under a Laplacian model of the quantized coefficients. This objective is solved by accelerated proximal gradient descent, and unrolling the iterations produces a nonlinear feed-forward encoder network whose layers are interpretable as gradient steps and soft-thresholding operations. The decoder requires the inverse and inverse square root of the Gram matrix ΦᵀΦ, which are approximated by unrolled conjugate gradient descent and a truncated Taylor series, avoiding the O(N³) eigendecomposition of prior work. A prediction module predicts high-pass coefficients from l

Load-bearing premise

The derivation of the RD-optimal encoder assumes that the quantized transform coefficients are independent, zero-mean Laplacian random variables with a common scale per resolution level, which justifies the L1 penalty; if the actual coefficients do not follow this model, the claimed optimality of the unrolled PGD encoder is not established.

Editorial extensions

If this is right

  • If the claims hold, G-PCC-style attribute coding can be improved with a fully interpretable, unrolled network that uses fewer than 150 trainable parameters per resolution level.
  • Higher-order B-spline subspaces (p=2) yield continuous attribute reconstructions and cut bit rate by more than 20% over piecewise-constant RAHT when no prediction is used.
  • Encoding complexity becomes a tunable knob: increasing the number of unrolled PGD iterations recovers rate-distortion performance when the decoder is fixed at low complexity.
  • The same volumetric framework, with different linear functionals, extends to coding radiance fields as well as point cloud attributes, since the distortion norm can be defined through ray integrals.
  • The trained predictor and subspace-defining kernels adapt to data, which is the main source of the reported gains over the handcrafted G-PCC predictor.

Reading between the lines

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

  • Editorial inference: the paper does not empirically compare the learned coefficients' histograms to the assumed Laplacian model; if the mismatch is large, the 'RD-optimal' encoder claim is formally weakened, though the learned thresholds and scales might still yield practical gains.
  • Editorial inference: the same unrolled PGD encoder could be used with content-adaptive fine-tuning at encode time, potentially improving the rate-distortion gains beyond the static trained models reported here.
  • Editorial inference: the variable-complexity encoder suggests an application-driven compute budget: a mobile encoder could use few PGD iterations while a cloud encoder uses many, with graceful rate-distortion degradation.
  • Editorial inference: the radiance-field extension noted in the paper could be tested by instantiating the ray-based linear functionals and learning B-spline subspaces for density and color fields, which would connect this work to neural radiance field compression.
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

3 major / 6 minor

Summary. This paper proposes a learned 3D point cloud attribute codec within a multi-resolution B-spline projection framework. The encoder is an unrolled proximal gradient descent (PGD) algorithm that minimizes a rate-distortion Lagrangian in which the rate term is an L1 penalty derived from a per-level zero-mean Laplacian model of quantized transform coefficients; the decoder uses unrolled conjugate gradient descent and truncated Taylor expansions to approximate Gram-matrix inversions, yielding O(M1 M2 M3 N) complexity. A trainable coarse-to-fine prediction module is integrated into the transform. Experiments on four MPEG point clouds report a 6-11% bitrate reduction and about 1 dB PSNR gain relative to an estimated G-PCC baseline using RAHT(p=1) with prediction.

Significance. If the empirical claims are confirmed, this is a valuable contribution: it provides a practical, trainable generalization of RAHT with interpretable unrolled optimization, few trainable parameters per level, variable encoder/decoder complexity, and a principled rate-distortion objective. The derivations of the PGD encoder and the complexity-reducing approximations are clear and largely self-contained. The main uncertainty is not the internal mathematics but whether the headline gain survives comparison with the actual MPEG G-PCC codec and whether the Laplacian rate model matches the true entropy of the learned coefficients.

major comments (3)
  1. [VI.B, VI.D, Eq. (87)] The headline 6-11% bitrate reduction is measured against a baseline that the authors estimate, not against the actual G-PCC/TMC13 codec. The baseline uses RLGR entropy coding and a specific implementation of the MPEG predictor via Eq. (87). G-PCC's RAHT predictor, arithmetic coder, and rate control may differ in ways that change the comparison by several percent. Please report results against actual G-PCC with RAHT(p=1)+prediction, or provide the exact baseline code/bitstreams so the comparison can be reproduced.
  2. [IV.C, Eqs. (48)-(56); VI.A] The PGD encoder is RD-optimal only for the assumed i.i.d. zero-mean Laplacian coefficient model. The model parameters (m_l, b_l, Delta) are learned, but the paper does not show that the actual coefficients produced by the learned transform are Laplacian, nor that the Laplacian entropy proxy matches the RLGR rate used in evaluation. In addition, the relaxed real-valued objective is quantized open-loop (Sec. V), so the encoder does not minimize the integer-quantized RD objective. Please validate the Laplacian assumption empirically (e.g., histograms vs. fitted Laplacians) or demonstrate that the mismatch does not affect the RD results.
  3. [VI.B] The training and evaluation protocol is not reproducible: the training set is described only as 'approximately more than 1000 point clouds' from Sketchfab, with no list of models, crop-generation details, or train/test split, and no code is provided. Since the empirical claim is central, please release the data-generation code and model code, or provide a sufficiently detailed protocol.
minor comments (6)
  1. [Eq. (48)] The Laplacian density is written as p(x)=1/(2b) exp(|x|/b); the exponent should be -|x|/b.
  2. [Eq. (76)] The vector V is written as [F_l0; G_l0; ...; G_L], but the hierarchy uses coefficients up to G_{L-1}; please reconcile the indexing.
  3. [Table I] The entry 'M1=5, M1=15 for RAHT(1) and RAHT(2)' appears to duplicate M1; presumably one of these should refer to M2 (Taylor order) or the text should clarify.
  4. [Sec. VI.E] The sentence 'flexibility to shift complexity between encoder and encoder' should read 'between encoder and decoder.'
  5. [Eq. (75)] There is an unbalanced parenthesis in the last expression: (Φ_L^T Θ)V^{(t)} has an extra closing parenthesis.
  6. [Sec. VI.B] The text says 'six levels of resolution from l0=4 to L=10'; that range contains seven levels if endpoints are included. Please clarify the count.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core derivation and empirical comparison are self-contained; self-citations are not load-bearing.

full rationale

The derivation chain is self-contained. The RD objective (Eq. 47) is defined directly as D + λR; the L1 rate term is derived from an explicit Laplacian model (Eqs. 48–56) rather than from the target result. The unrolled PGD encoder (Eqs. 70–72) is constructed from that objective, and its parameters are trained end-to-end (Sec. VI.A). The reported bitrates are measured by RLGR entropy coding (Sec. VI.B), which is independent of the trained Laplacian proxy, so the empirical gain is not obtained by minimizing the very metric that is later reported. The G-PCC baseline is described as an 'estimate' (Sec. VI.B) using a fixed predictor formula (Eq. 87) and RAHT(p=1); whether or not that estimate accurately matches the true G-PCC is a correctness/reproducibility concern, not a circular reduction, since the baseline is not defined in terms of the paper's learned model. Self-citations to [1], [32]–[34] point to prior published work, but the paper rederives the framework in Section III and does not invoke a uniqueness theorem or any unverified premise from those papers. No equation reduces to another by construction beyond the standard linear-algebra identities explicitly derived. The Laplacian assumption and the estimated baseline are limitations, but they do not render any claimed prediction equivalent to its input by definition.

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

The central derivation rests on standard Hilbert space and B-spline refinement results, plus an assumed Laplacian coefficient model whose parameters are then learned from data. No new physical entities are introduced.

free parameters (5)
  • CGD step sizes and momenta [α_t, β_t] per level = trained; initialized α_t=0.99, β_t=0.5 for all t
    Unrolled conjugate gradient iterations approximate the inverse of the Gram matrix (Φ_l^T Φ_l); these are learnable per-iteration parameters.
  • Taylor coefficients for X^{-1/2} = trained; initialized from Taylor expansion coefficients of 1/sqrt(x)
    Used to approximate orthonormalization operators R_Φl and R_Ψl.
  • PGD step sizes, momentum, shrinkage [α_t, β_t, γ̃_t^l] = trained; initialized α_t=0.8, β_t=0.1, γ̃=2e-5 or 2e-3
    Parameters of the unrolled proximal gradient descent encoder.
  • Laplace distribution location, scale, quantization stepsize [m_l, b_l, Δ] = trained; initialized m_l=0 or 1, b_l=0.001
    Entropy model used in the training loss and in deriving the L1 rate penalty.
  • Prediction kernels P_l and transform kernels A_l = trained; initialized from RAHT(1) or RAHT(2)
    Convolutional weights for refinement and prediction across levels.
assumptions (5)
  • standard math F with inner product (9) is a Hilbert space after quotient by equivalence classes
    Section III-C: needed for orthogonal projection and the transform coding framework.
  • standard math B-spline refinement coefficients a_k in (17) are convolutional and geometry-independent
    Section III-F: used to propagate analysis coefficients between levels.
  • domain assumption At finest level L, each B-spline basis function contains exactly one point cloud point and evaluates to 1 there, so Φ_L^T Φ_L = I
    Section III-F: initializes the analysis ladder with the point attribute values.
  • ad hoc to paper Quantized transform coefficients are i.i.d. zero-mean Laplacian per level
    Section IV-C equations (48)-(56): this model underlies the derivation of the L1 rate penalty; actual coefficient statistics are not verified.
  • standard math PGD and the Taylor/CGD approximations converge with the chosen stepsizes; μ small enough
    Appendix VIII-A to VIII-C: convergence conditions are standard but assumed to hold for the learned parameters.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Unrolling of Sparsity-Induced RDO for 3D Point Cloud Attribute Coding." pith.science (2026). https://pith.science/paper/AOWXPWI7

@misc{pith2026250908685,
  author       = {Pith},
  title        = {Pith review of: Deep Unrolling of Sparsity-Induced RDO for 3D Point Cloud Attribute Coding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AOWXPWI7}},
  note         = {Machine review of arXiv:2509.08685}
}
abstract

Given encoded 3D point cloud geometry available at the decoder, we study the problem of lossy attribute compression in a multi-resolution B-spline projection framework. A target continuous 3D attribute function is first projected onto a sequence of nested subspaces $\mathcal{F}^{(p)}_{l_0} \subseteq \cdots \subseteq \mathcal{F}^{(p)}_{L}$, where $\mathcal{F}^{(p)}_{l}$ is a family of functions spanned by a B-spline basis function of order $p$ at a chosen scale and its integer shifts. The projected low-pass coefficients $F_l^*$ are computed by variable-complexity unrolling of a rate-distortion (RD) optimization algorithm into a feed-forward network, where the rate term is the sparsity-promoting $\ell_1$-norm. Thus, the projection operation is end-to-end differentiable. For a chosen coarse-to-fine predictor, the coefficients are then adjusted to account for the prediction from a lower-resolution to a higher-resolution, which is also optimized in a data-driven manner.

Figures

Figures reproduced from arXiv: 2509.08685 by the authors.

Figure 1
Figure 1. Ladder structure of encoder and decoder using orthonor￾mal transform, analogous to a U-Net architecture, with quanti￾zation in between. The orange (and yellow) nodes correspond to the inputs (and outputs) of the blocks [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Sub-networks for Zl, Z ⊤ l , and (Φ⊤Φ) −1 . We describe Conjugate Gradient Descent Unrolling in Appendix VIII-B. cloud standard with RAHT(p = 1). In that case, both (Φ⊤ l Φl) and (Ψ⊤ l Ψl) are diagonal, in which case their inverses and inverse square roots are also diagonal, and the operators are O(N). The situation is fundamentally due to the following well-known property of wavelets: the only orthogonal wavelets w… view at source ↗
Figure 3
Figure 3. M2-layer network approximating the ortho￾normalization operator RX = X−1/2 , where the block X can be either the O(N) operator Φ⊤ l Φl or the O(M1N) subnetwork approximating the operator Ψ⊤ l Ψl. by unrolling it into an outer loop with M2 steps and an inner loop with M1 steps, for an overall complexity of O(M1M2N). In particular, (Ψ⊤ l Ψl) −1/2 can be im￾plemented by a feedforward network with M2 layers of feedforwa… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Encoder as unrolled rate-distortion optimization. where Proxτ (u) ∆= arg min v  1 2τ (v − u) 2 + |v|  (68) =    u − τ if u > τ u + τ if u < −τ 0 if |u| ≤ τ (69) is a proximal operator, also known as a shrinkage operator in this case. The shrinkage operator pulls i…
Figure 5
Figure 5. Figure 5: Encoder-decoder pipeline with parameters Ω trained end-to-end. where ΘP = ΘM and Θ⊤ P = M⊤Θ⊤. Note that while M⊤ may be inconvenient to compute, it is shown in Appendix VIII-G that M−1 may be used instead of M⊤, possibly at the expense of slower convergence. After the …
Figure 6
Figure 6. Figure 6: Rate-Distortion curves: (a) Longdress, (b) Redandblack, (c) Loot, (d) Soldier complexity. In all experiments, we initialize the (3×3× 3) kernels of the operations Al , A⊤ l , and Φ⊤ l Φl in two different ways, from the kernels for RAHT(p = 1) and RAHT(p = 2), respectiv…
Figure 7
Figure 7. Figure 7: Reconstructed Point Cloud in case without prediction at ∼ 0.529 bits per point: (left) RAHT(p=1) with 30.85dB, (right) RAHT(p=2) with 31.84dB [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Rate-Distortion curves: (left) Loot, (right) Soldier The reduced bit rate is at least 20% across all test points clouds. This is expected because functions in the subspace spanned by the higher-order B-splines are con￾tinuous and piecewise linear, guaranteeing a contin…
Figure 9
Figure 9. Figure 9: Complexity analysis curves Redandblack and also the ability to learn the kernels for Al that define the sequence of subspaces. We also observed that the performance of the learned models initialized at p = 2 is not as good as that of models initialized at p = 1. We bel…
Figure 10
Figure 10. Figure 10: Taylor series for x −1/2 around x = a = 1/(2µ). F. Change of basis We first give an example of the transformation M(3) for a simple case of 3 levels, F1 = F0 ⊕ G0 ⊕ G1   G1 G0 F0   = M(3)    G ′′ 1 G ′′ 0 F0    (133) where the last 2 rows of M(3) is already …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 1 canonical work pages

  1. [1]

    A volumetric approach to point cloud compression—part i: Attribute compres- sion,

    P. A. Chou, M. Koroteev, and M. Krivoku ´ca, “A volumetric approach to point cloud compression—part i: Attribute compres- sion,”IEEE Transactions on Image Processing, vol. 29, pp. 2203– 2216, 2020

    2020

  2. [2]

    Compression of 3d point clouds using a region-adaptive hierarchical transform,

    R. L. de Queiroz and P. A. Chou, “Compression of 3d point clouds using a region-adaptive hierarchical transform,”IEEE Transactions on Image Processing, vol. 25, no. 8, pp. 3947–3956, 2016

    2016

  3. [3]

    Integer alternative for the region-adaptive hierarchical trans- form,

    G. P. Sandri, P. A. Chou, M. Krivoku ´ca, and R. L. de Queiroz, “Integer alternative for the region-adaptive hierarchical trans- form,”IEEE Signal Processing Letters, vol. 26, no. 9, pp. 1369– 1372, 2019

    2019

  4. [4]

    Emerging MPEG standards for point cloud compression,

    S. Schwarz, M. Preda, V . Baroncini, M. Budagavi, P. Cesar, P. A. Chou, R. A. Cohen, M. Krivoku ´ca, S. Lasserre, Z. Li, J. Llach, K. Mammou, R. Mekuria, O. Nakagami, E. Siahaan, A. Tabatabai, A. Tourapis, and V . Zakharchenko, “Emerging MPEG standards for point cloud compression,”IEEE J. Emerging Topics in Circuits and Systems, vol. 9, no. 1, pp. 133–148...

    2019

  5. [5]

    Saad,Iterative methods for sparse linear systems

    Y . Saad,Iterative methods for sparse linear systems. SIAM, 2003

    2003

  6. [6]

    Proximal algorithms,

    N. Parikh and S. Boyd, “Proximal algorithms,” inFoundations and Trends in Optimization, vol. 1, no.3, 2013, pp. 123–231

    2013

  7. [7]

    Nonlinear transform coding,

    J. Ball ´e, P. A. Chou, D. Minnen, S. Singh, N. Johnston, E. Agusts- son, S. J. Hwang, and G. Toderici, “Nonlinear transform coding,” IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 2, pp. 339–353, 2021

    2021

  8. [8]

    On an improvement of RAHT to ex- ploit attribute correlation,

    S. Lasserre and D. Flynn, “On an improvement of RAHT to ex- ploit attribute correlation,” ISO/IEC MPEG JTC1/SC29/WG11, Geneva, CH, input document m47378, Mar. 2019

    2019

Show all 54 references
  1. [9]

    Point cloud attribute compression with graph transform,

    C. Zhang, D. Flor ˆencio, and C. Loop, “Point cloud attribute compression with graph transform,” in2014 IEEE International Conference on Image Processing (ICIP), 2014, pp. 2066–2070

    2014

  2. [10]

    Attribute compression for sparse point clouds using graph transforms,

    R. A. Cohen, D. Tian, and A. Vetro, “Attribute compression for sparse point clouds using graph transforms,” in2016 IEEE International Conference on Image Processing (ICIP), 2016, pp. 1374–1378

    2016

  3. [11]

    Spectral folding and two-channel filter-banks on arbitrary graphs,

    E. Pavez, B. Girault, A. Ortega, and P. A. Chou, “Spectral folding and two-channel filter-banks on arbitrary graphs,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 5070–5074

    2021

  4. [12]

    Ray tracing volume densities,

    J. T. Kajiya and B. P. V on Herzen, “Ray tracing volume densities,” inProceedings of the 11th Annual Conference on Computer Graphics and Interactive Techniques, ser. SIGGRAPH ’84. New York, NY , USA: Association for Computing Machinery, 1984, p. 165–174. [Online]. Available: h...

    1984

  5. [13]

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

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” inComputer Vision – ECCV 2020, A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, Eds. Cham: Springer International P...

    2020

  6. [14]

    Ewa splatting,

    M. Zwicker, H. Pfister, J. Van Baar, and M. Gross, “Ewa splatting,”IEEE Transactions on Visualization and Computer Graphics, vol. 8, no. 3, pp. 223–238, 2002

    2002

  7. [15]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  8. [16]

    Com- pressed 3d gaussian splatting for accelerated novel view synthe- sis,

    S. Niedermayr, J. Stumpfegger, and R. Westermann, “Com- pressed 3d gaussian splatting for accelerated novel view synthe- sis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 349–10 358

    2024

  9. [17]

    Mesongs: Post-training compression of 3d gaussians via efficient attribute transformation,

    S. Xie, W. Zhang, C. Tang, Y . Bai, R. Lu, S. Ge, and Z. Wang, “Mesongs: Post-training compression of 3d gaussians via efficient attribute transformation,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 434–452

    2024

  10. [18]

    Hac: Hash- grid assisted context for 3d gaussian splatting compression,

    Y . Chen, Q. Wu, W. Lin, M. Harandi, and J. Cai, “Hac: Hash- grid assisted context for 3d gaussian splatting compression,” in European Conference on Computer Vision. Springer, 2024, pp. 422–438. 13

    2024

  11. [19]

    Compgs: Efficient 3d scene representation via compressed gaussian splatting,

    X. Liu, X. Wu, P. Zhang, S. Wang, Z. Li, and S. Kwong, “Compgs: Efficient 3d scene representation via compressed gaussian splatting,” inProceedings of the 32nd ACM International Conference on Multimedia, ser. MM ’24. New York, NY , USA: Association for Computing Machinery, 202...

    2024

  12. [20]

    Deep learning-based point cloud geometry coding: RD control through implicit and explicit quantization,

    A. F. R. Guarda, N. M. M. Rodrigues, and F. Pereira, “Deep learning-based point cloud geometry coding: RD control through implicit and explicit quantization,” in2020 IEEE Int. Conf. on Multimedia & Expo Wksps. (ICMEW), 2020

    2020

  13. [21]

    Deep implicit volume compression,

    D. Tang, S. Singh, P. A. Chou, C. H ¨ane, M. Dou, S. Fanello, J. Taylor, P. Davidson, O. G. Guleryuz, Y . Zhang, S. Izadi, A. Tagliasacchi, S. Bouaziz, and C. Keskin, “Deep implicit volume compression,” in2020 IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  14. [22]

    Towards neural network approaches for point cloud compression,

    E. Alexiou, K. Tung, and T. Ebrahimi, “Towards neural network approaches for point cloud compression,” inApplications of Digital Image Processing XLIII, A. G. Tescher and T. Ebrahimi, Eds., vol. 11510, International Society for Optics and Photonics. SPIE, 2020, p. 1151008. [On...

    2020 doi

  15. [23]

    Lvac: Learned volumetric attribute compression for point clouds using coordinate based networks,

    B. Isik, P. A. Chou, S. J. Hwang, N. Johnston, and G. Toderici, “Lvac: Learned volumetric attribute compression for point clouds using coordinate based networks,”Frontiers in Signal Processing, vol. 2, 2022. [Online]. Available: https://www.frontiersin.org/ journals/signal-pro...

    2022

  16. [24]

    Sparse tensor-based point cloud attribute compression,

    J. Wang and Z. Ma, “Sparse tensor-based point cloud attribute compression,” in2022 IEEE Int’l Conf. Multimedia Information Processing and Retrieval (MIPR), 2022

    2022

  17. [25]

    Scalable point cloud at- tribute compression,

    J. Zhang, J. Wang, D. Ding, and Z. Ma, “Scalable point cloud at- tribute compression,”IEEE Transactions on Multimedia, vol. 27, pp. 889–899, 2025

    2025

  18. [26]

    Entropy-constrained vector quantization,

    P. Chou, T. Lookabaugh, and R. Gray, “Entropy-constrained vector quantization,”IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 37, no. 1, pp. 31–42, 1989

    1989

  19. [27]

    Variational image compression with a scale hyperprior,

    J. Ball ´e, D. Minnen, S. Singh, S. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” 01 2018

    2018

  20. [28]

    Algorithm unrolling: Inter- pretable, efficient deep learning for signal and image processing,

    V . Monga, Y . Li, and Y . C. Eldar, “Algorithm unrolling: Inter- pretable, efficient deep learning for signal and image processing,” IEEE Signal Processing Magazine, vol. 38, no. 2, pp. 18–44, 2021

    2021

  21. [29]

    Learning fast approximations of sparse coding,

    K. Gregor and Y . LeCun, “Learning fast approximations of sparse coding,” inProceedings of the 27th International Conference on International Conference on Machine Learning, ser. ICML’10. Madison, WI, USA: Omnipress, 2010, p. 399–406

    2010

  22. [30]

    White-box transformers via sparse rate reduction,

    Y . Yu, S. Buchanan, D. Pai, T. Chu, Z. Wu, S. Tong, B. Haeffele, and Y . Ma, “White-box transformers via sparse rate reduction,” inAdvances in Neural Information Processing Systems, A. Oh, T. Neumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Ass...

    2023

  23. [31]

    Interpretable lightweight transformer via unrolling of learned graph smoothness priors,

    V . H. T. T. DO, P. Eftekhar, S. A. Hosseini, G. Cheung, and P. Chou, “Interpretable lightweight transformer via unrolling of learned graph smoothness priors,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. [Online]. Available: https://op...

    2024

  24. [32]

    V olumetric attribute compression for 3d point clouds using feedforward network with geometric attention,

    T. T. Do, P. A. Chou, and G. Cheung, “V olumetric attribute compression for 3d point clouds using feedforward network with geometric attention,” inICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  25. [33]

    Learned nonlinear predictor for critically sampled 3d point cloud attribute compression,

    ——, “Learned nonlinear predictor for critically sampled 3d point cloud attribute compression,” in2024 IEEE International Conference on Image Processing (ICIP). IEEE, 2024, pp. 3299– 3305

    2024

  26. [34]

    V olumetric 3d point cloud attribute compression: Learned polynomial bilateral filter for prediction,

    ——, “V olumetric 3d point cloud attribute compression: Learned polynomial bilateral filter for prediction,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 3915–3919

    2024

  27. [35]

    Equivalence class — Wikipedia, the free encyclopedia,

    Wikipedia contributors, “Equivalence class — Wikipedia, the free encyclopedia,” https://en.wikipedia.org/w/index.php?title= Equivalence class&oldid=1304551778, 2025

    2025

  28. [36]

    Hilbert space — Wikipedia, the free encyclope- dia,

    ——, “Hilbert space — Wikipedia, the free encyclope- dia,” https://en.wikipedia.org/w/index.php?title=Hilbert space& oldid=1303408008, 2025

    2025

  29. [37]

    G-PCC codec description v12,

    3DG, “G-PCC codec description v12,” ISO/IEC MPEG JTC1/SC29/WG11, Approved WG 11 document N00151, Septemper 2021

    2021

  30. [38]

    Gram matrix — Wikipedia, the free encyclopedia,

    Wikipedia contributors, “Gram matrix — Wikipedia, the free encyclopedia,” https://en.wikipedia.org/w/index.php?title=Gram matrix&oldid=1300063302, 2025

    2025

  31. [39]

    Daubechies,Ten Lectures on Wavelets

    I. Daubechies,Ten Lectures on Wavelets. Society for Industrial and Applied Mathematics, 1992. [Online]. Available: https://epubs.siam.org/doi/abs/10.1137/1.9781611970104

    1992 doi

  32. [40]

    Vetterli and J

    M. Vetterli and J. Kovacevi ´c,Wavelets and Subband Coding. Prentice Hall, 1995

    1995

  33. [41]

    Conjugate gradient method — Wikipedia, the free encyclopedia,

    Wikipedia contributors, “Conjugate gradient method — Wikipedia, the free encyclopedia,” 2025, [Online; accessed 25- February-2025]. [Online]. Available: https://en.wikipedia.org/w/ index.php?title=Conjugate gradient method&oldid=1276055255

    2025

  34. [42]

    Deepsdf: Learning continuous signed distance functions for shape representation,

    J. J. Park, P. Florence, J. Straub, R. Newcombe, and S. Lovegrove, “Deepsdf: Learning continuous signed distance functions for shape representation,” inThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  35. [43]

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

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” inECCV, 2020

    2020

  36. [44]

    Lvac: Learned volumetric attribute compression for point clouds using coordinate based networks,

    B. Isik, P. A. Chou, S. J. Hwang, N. Johnston, and G. Toderici, “Lvac: Learned volumetric attribute compression for point clouds using coordinate based networks,”arXiv preprint arXiv:2111.08988, 2021

    2021 arXiv

  37. [45]

    Coin: Compression with implicit neural representations,

    E. Dupont, A. Goli ´nski, M. Alizadeh, Y . W. Teh, and A. Doucet, “Coin: Compression with implicit neural representations,” ArXiv, vol. abs/2103.03123, 2021. [Online]. Available: https: //api.semanticscholar.org/CorpusID:232110691

    2021 arXiv

  38. [46]

    D’oh: Decoder-only random hypernetworks for implicit neural representations,

    C. Gordon, L. E. MacDonald, H. Saratchandran, and S. Lucey, “D’oh: Decoder-only random hypernetworks for implicit neural representations,” inComputer Vision – ACCV 2024, M. Cho, I. Laptev, D. Tran, A. Yao, and H. Zha, Eds. Singapore: Springer Nature Singapore, 2025, pp. 128–147

    2024

  39. [47]

    Proximal algorithms,

    N. Parikh and S. Boyd, “Proximal algorithms,”Found. Trends Optim., vol. 1, no. 3, p. 127–239, Jan. 2014. [Online]. Available: https://doi.org/10.1561/2400000003

    2014 doi

  40. [48]

    T. M. Cover and J. A. Thomas,Elements of Information Theory. John Wiley & Sons, 2006

    2006

  41. [49]

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems,

    A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding algorithm for linear inverse problems,”SIAM journal on imaging sciences, vol. 2, no. 1, pp. 183–202, 2009

    2009

  42. [50]

    Multi- resolution intra-predictive coding of 3d point cloud attributes,

    E. Pavez, A. L. Souto, R. L. D. Queiroz, and A. Ortega, “Multi- resolution intra-predictive coding of 3d point cloud attributes,” in2021 IEEE International Conference on Image Processing (ICIP), 2021, pp. 3393–3397

    2021

  43. [51]

    Adaptive run-length/golomb-rice encoding of quan- tized generalized gaussian sources with unknown statistics,

    H. Malvar, “Adaptive run-length/golomb-rice encoding of quan- tized generalized gaussian sources with unknown statistics,” in Data Compression Conference (DCC’06), 2006, pp. 23–32

    2006

  44. [52]

    8i vox- elized full bodies — a voxelized point cloud dataset,

    E. d’Eon, B. Harrison, T. Meyers, and P. A. Chou, “8i vox- elized full bodies — a voxelized point cloud dataset,” ISO/IEC JTC1/SC29 WG1 & WG11 (JPEG & MPEG), Ljubljana, Slove- nia, input document M74006 & m42914, Jan. 2017

    2017

  45. [53]

    Peer upsampled transform domain prediction for g-pcc,

    W. Wang, Y . Xu, K. Zhang, and L. Zhang, “Peer upsampled transform domain prediction for g-pcc,” in2023 IEEE Interna- tional Conference on Multimedia and Expo (ICME), 2023, pp. 708–713

    2023

  46. [54]

    1 + x0 −x x0 + x0 −x x0 2 +· · · # = 1 x0

    Y . Nesterov,Introductory lectures on convex optimization: A basic course. Springer Science & Business Media, 2013, vol. 87. 14 VIII. APPENDIX A. Taylor expansion ofX −1 and gradient descent The Taylor expansion ofX −1 totterms is equivalent to the solution of a minimum-norm p...

    2013

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.