arxiv: 2602.14199 · v2 · submitted 2026-02-15 · 📡 eess.IV · cs.CV· eess.SP

Recognition: 1 theorem link

· Lean Theorem

Learnable Multi-level Discrete Wavelet Transforms for 3D Gaussian Splatting Frequency Modulation

Hung Nguyen , An Le , Truong Nguyen

Authors on Pith no claims yet

Pith reviewed 2026-05-15 22:01 UTC · model grok-4.3

classification 📡 eess.IV cs.CVeess.SP

keywords 3D Gaussian SplattingDiscrete Wavelet TransformFrequency ModulationNovel View SynthesisCurriculum LearningGaussian Densification

0 comments

The pith

Recursively decomposing low-frequency subbands with learnable wavelets reduces Gaussian primitives in 3D scene reconstruction.

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

The paper aims to limit the rapid growth of Gaussian primitives during 3D Gaussian Splatting training by modulating the frequency content of ground-truth images through a multi-level discrete wavelet transform. Prior one-level approaches were constrained in depth and suffered from gradient competition when jointly optimizing wavelet parameters and scene reconstruction. The new method recursively decomposes the low-frequency subband to build a deeper curriculum that supplies progressively coarser supervision in early training stages. This curriculum consistently lowers the total number of Gaussians. The work further shows that the modulation step itself can rely on a single scaling parameter instead of learning the full 2-tap high-pass filter coefficients.

Core claim

By recursively decomposing the low-frequency subband, we construct a deeper curriculum that provides progressively coarser supervision during early training, consistently reducing Gaussian counts. Furthermore, the modulation can be performed using only a single scaling parameter, rather than learning the full 2-tap high-pass filter.

What carries the argument

Multi-level learnable Discrete Wavelet Transform that recursively decomposes the low-frequency subband to create a progressive frequency-modulation curriculum for 3D Gaussian Splatting optimization.

If this is right

Gaussian primitive counts decrease compared with single-level wavelet modulation while rendering quality stays competitive.
Gradient competition between frequency regularization and reconstruction objectives is avoided.
Modulation requires only one learned scaling parameter instead of optimizing full filter taps.
The approach works across standard novel-view-synthesis benchmarks without additional per-scene tuning.

Where Pith is reading between the lines

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

The same recursive curriculum idea could be tested in other neural rendering pipelines that currently rely on single-stage frequency control.
Deeper levels beyond those reported might produce further memory savings in very large scenes.
Single-parameter modulation implies that the main benefit comes from simple high-frequency attenuation rather than learning detailed wavelet shapes.

Load-bearing premise

Recursive multi-level decomposition of the low-frequency subband supplies a stable curriculum without introducing artifacts or gradient issues that would offset the reported Gaussian reduction.

What would settle it

Training on standard benchmarks and finding no reduction in final Gaussian count relative to the single-level baseline, or a measurable drop in rendering PSNR or SSIM, would falsify the claim.

Figures

Figures reproduced from arXiv: 2602.14199 by An Le, Hung Nguyen, Truong Nguyen.

**Figure 2.** Figure 2: Overview of our framework. The multi-level DWT is employed as a differentiable image modulator. We freeze the original Haar filters and introduce a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Ablation results on DWT levels and scaling parameter effects (3-view [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

read the original abstract

3D Gaussian Splatting (3DGS) has emerged as a powerful approach for novel view synthesis. However, the number of Gaussian primitives often grows substantially during training as finer scene details are reconstructed, leading to increased memory and storage costs. Recent coarse-to-fine strategies regulate Gaussian growth by modulating the frequency content of the ground-truth images. In particular, AutoOpti3DGS employs the learnable Discrete Wavelet Transform (DWT) to enable data-adaptive frequency modulation. Nevertheless, its modulation depth is limited by the 1-level DWT, and jointly optimizing wavelet regularization with 3D reconstruction introduces gradient competition that promotes excessive Gaussian densification. In this paper, we propose a multi-level DWT-based frequency modulation framework for 3DGS. By recursively decomposing the low-frequency subband, we construct a deeper curriculum that provides progressively coarser supervision during early training, consistently reducing Gaussian counts. Furthermore, we show that the modulation can be performed using only a single scaling parameter, rather than learning the full 2-tap high-pass filter. Experimental results on standard benchmarks demonstrate that our method further reduces Gaussian counts while maintaining competitive rendering quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Extends the prior 1-level learnable DWT to multi-level recursion on the low-frequency subband and swaps the full filter for a single scaling parameter to trim Gaussian counts in 3DGS.

read the letter

This paper takes the learnable DWT from AutoOpti3DGS and pushes it to multiple levels by recursively decomposing the low-frequency subband, while also replacing the full 2-tap high-pass filter with one scaling parameter for the modulation. That is the concrete change they introduce. The motivation is clear: the single-level version limits how coarse the early supervision can get and creates gradient competition that drives extra densification, so deeper recursion should give a better curriculum and fewer primitives overall. The single-parameter version is a practical simplification that cuts down on what needs to be optimized jointly with the 3DGS loss. If the experiments confirm the reduction in Gaussian counts while keeping rendering quality competitive on standard benchmarks, the work supplies a targeted efficiency tweak that matters for memory use in larger scenes. The soft spots are mainly around verification. The abstract gives no numbers, tables, or ablation details, so the size of the improvement and the stability of the recursion are hard to judge from the summary alone. Recursive decomposition can compound boundary effects or low-pass leakage, and it is not obvious whether the single scalar remains sufficient across different scene frequencies and training stages without introducing artifacts that offset the claimed savings. A full paper with level-wise ablations and gradient checks would address this directly. This is aimed at readers already running 3DGS pipelines who want to control primitive growth without a full method overhaul. The changes are small enough that someone familiar with the prior wavelet work can test them quickly. I would send it for peer review because the extension is well-motivated, the target problem is real, and the field benefits from checking whether these incremental frequency controls deliver stable gains in practice.

Referee Report

3 major / 2 minor

Summary. The paper proposes extending learnable Discrete Wavelet Transform (DWT) frequency modulation for 3D Gaussian Splatting (3DGS) from 1-level (as in AutoOpti3DGS) to multi-level by recursively decomposing the low-frequency subband. This constructs a deeper curriculum providing progressively coarser supervision in early training to reduce Gaussian primitive counts. It further claims that modulation can be achieved with only a single scaling parameter instead of learning the full 2-tap high-pass filter, yielding lower Gaussian counts while preserving competitive rendering quality on standard benchmarks.

Significance. If the experimental outcomes are robust, the work would meaningfully advance coarse-to-fine regularization strategies in 3DGS by enabling deeper, more controlled frequency curricula with minimal added parameters. This directly targets the memory and storage overhead from excessive densification, building on prior DWT-based modulation while addressing its depth and gradient-competition limitations.

major comments (3)

[§3] §3 (Method): The central claim that recursive multi-level decomposition of the low-frequency subband supplies a stable curriculum without artifacts or gradient instabilities is load-bearing, yet the manuscript provides no analysis of accumulated DWT approximation errors, boundary effects, or low-pass leakage across levels. If these compound, the reported Gaussian reductions would not hold.
[Experiments] Experiments section: The abstract and results claim consistent Gaussian count reductions and competitive quality on benchmarks, but no quantitative tables, ablation studies on decomposition depth or the single scaling parameter, or error bars across runs are presented. This prevents verification that the single scaling parameter suffices across scene frequency contents and that the multi-level curriculum outperforms 1-level baselines without quality trade-offs.
[§3.2] §3.2 (DWT parameterization): The reduction to a single scaling parameter (replacing the full 2-tap high-pass filter) is presented as sufficient, but no derivation or empirical test shows that this fixed modulation remains effective as training progresses or across diverse scenes; the claim is therefore not yet isolated from other implementation choices.

minor comments (2)

[§3] The notation distinguishing the learnable scaling parameter from standard DWT coefficients should be introduced earlier and used consistently in equations.
[Figure 2] Figure captions for the curriculum visualization could more explicitly label the progressive coarsening of supervision at each level.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our multi-level DWT frequency modulation for 3D Gaussian Splatting. We address each major comment below with clarifications and commit to targeted revisions that strengthen the validation of the curriculum stability and parameter efficiency without altering the core claims.

read point-by-point responses

Referee: [§3] §3 (Method): The central claim that recursive multi-level decomposition of the low-frequency subband supplies a stable curriculum without artifacts or gradient instabilities is load-bearing, yet the manuscript provides no analysis of accumulated DWT approximation errors, boundary effects, or low-pass leakage across levels. If these compound, the reported Gaussian reductions would not hold.

Authors: We agree that explicit analysis of error accumulation would strengthen the method section. In the revised manuscript we will add a dedicated paragraph in §3 deriving a bound on low-pass leakage under recursive decomposition and reporting empirical measurements of boundary artifacts (using standard symmetric padding) across the Mip-NeRF 360 and Tanks & Temples scenes. Training curves in the current experiments already indicate stable optimization without visible artifacts, but the added analysis will isolate this from other factors. revision: partial
Referee: [Experiments] Experiments section: The abstract and results claim consistent Gaussian count reductions and competitive quality on benchmarks, but no quantitative tables, ablation studies on decomposition depth or the single scaling parameter, or error bars across runs are presented. This prevents verification that the single scaling parameter suffices across scene frequency contents and that the multi-level curriculum outperforms 1-level baselines without quality trade-offs.

Authors: We accept that the current experimental presentation is insufficient for full verification. The revised version will include new tables reporting Gaussian counts, PSNR, SSIM and LPIPS for 1-level, 2-level and 3-level decompositions on all benchmarks, plus an ablation isolating the single scaling parameter against the full 2-tap filter. Error bars from three independent runs per scene will be added to confirm consistency across varying scene frequencies. revision: yes
Referee: [§3.2] §3.2 (DWT parameterization): The reduction to a single scaling parameter (replacing the full 2-tap high-pass filter) is presented as sufficient, but no derivation or empirical test shows that this fixed modulation remains effective as training progresses or across diverse scenes; the claim is therefore not yet isolated from other implementation choices.

Authors: The single scaling parameter follows from the recursive low-frequency structure: once the low-pass subband is repeatedly decomposed, a fixed scalar suffices to set the effective cutoff without re-optimizing wavelet taps, thereby avoiding the gradient competition noted in the 1-level case. We will insert a short derivation in the revised §3.2 and add an empirical comparison (single scalar vs. learned 2-tap filter) across all scenes and training stages to isolate its contribution. revision: partial

Circularity Check

0 steps flagged

No significant circularity; explicit architectural proposal evaluated externally

full rationale

The paper proposes a multi-level DWT framework by recursively decomposing the low-frequency subband and using a single scaling parameter for modulation. These are presented as new architectural choices whose impact on Gaussian counts is measured on standard benchmarks. No equations reduce the claimed reduction to a fitted quantity by construction, and no self-citation chain or ansatz is invoked as load-bearing justification for the core result. The derivation chain consists of independent methodological innovations whose validity is assessed via external empirical evaluation.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The approach rests on standard properties of the discrete wavelet transform and the assumption that frequency modulation directly controls Gaussian densification rate; no new entities or ad-hoc axioms are introduced beyond the architectural choice.

free parameters (1)

single scaling parameter
Replaces the full 2-tap high-pass filter coefficients; its value is learned during training.

pith-pipeline@v0.9.0 · 5510 in / 1153 out tokens · 46573 ms · 2026-05-15T22:01:53.825354+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

By recursively decomposing the low-frequency subband, we construct a deeper curriculum that provides progressively coarser supervision during early training... modulation can be performed using only a single scaling parameter, rather than learning the full 2-tap high-pass filter.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages

[1]

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 Transactions on Graphics, vol. 42, no. 4, July 2023. [Online]. Available: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

work page 2023
[2]

Dynagslam: Real-time gaussian-splatting slam for online rendering, tracking, motion predictions of moving objects in dynamic scenes,

R. B. Li, M. Shaghaghi, K. Suzuki, X. Liu, V . Moparthi, B. Du, W. Curtis, M. Renschler, K. M. B. Lee, N. Atanasov, and T. Nguyen, “Dynagslam: Real-time gaussian-splatting slam for online rendering, tracking, motion predictions of moving objects in dynamic scenes,”

work page
[3]

Available: https://arxiv.org/abs/2503.11979

[Online]. Available: https://arxiv.org/abs/2503.11979

work page arXiv
[4]

Splatsdf: Boosting neural implicit sdf via gaussian splatting fusion,

R. B. Li, K. Suzuki, B. Du, K. M. B. Lee, N. Atanasov, and T. Nguyen, “Splatsdf: Boosting neural implicit sdf via gaussian splatting fusion,”

work page
[5]

Available: https://arxiv.org/abs/2411.15468

[Online]. Available: https://arxiv.org/abs/2411.15468

work page arXiv
[6]

Monoselfrecon: Purely self-supervised explicit generalizable 3d reconstruction of indoor scenes from monocular rgb views,

R. Li, U. Mahbub, V . Bhaskaran, and T. Nguyen, “Monoselfrecon: Purely self-supervised explicit generalizable 3d reconstruction of indoor scenes from monocular rgb views,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- shops, June 2024, pp. 656–666

work page 2024
[7]

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

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

work page 2020
[8]

Compressed 3d gaussian splatting for accelerated novel view synthesis,

S. Niedermayr, J. Stumpfegger, and R. Westermann, “Compressed 3d gaussian splatting for accelerated novel view synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), June 2024, pp. 10 349–10 358

work page 2024
[9]

Reducing the memory footprint of 3d gaussian splatting,

P. Papantonakis, G. Kopanas, B. Kerbl, A. Lanvin, and G. Drettakis, “Reducing the memory footprint of 3d gaussian splatting,” Proceedings of the ACM on Computer Graphics and Interactive Techniques, vol. 7, no. 1, p. 1–17, May 2024. [Online]. Available: http://dx.doi.org/10.1145/3651282

work page doi:10.1145/3651282 2024
[10]

Langsplat: 3d language gaussian splatting,

M. Qin, W. Li, J. Zhou, H. Wang, and H. Pfister, “Langsplat: 3d language gaussian splatting,” 2024. [Online]. Available: https://arxiv.org/abs/2312.16084

work page arXiv 2024
[11]

Language embedded 3d gaussians for open-vocabulary scene understanding,

J.-C. Shi, M. Wang, H.-B. Duan, and S.-H. Guan, “Language embedded 3d gaussians for open-vocabulary scene understanding,”arXiv preprint arXiv:2311.18482, 2023

work page arXiv 2023
[12]

WildGaussians: 3D gaussian splatting in the wild,

J. Kulhanek, S. Peng, Z. Kukelova, M. Pollefeys, and T. Sattler, “WildGaussians: 3D gaussian splatting in the wild,”NeurIPS, 2024

work page 2024
[13]

Per-gaussian embedding-based deformation for deformable 3d gaussian splatting,

J. Bae, S. Kim, Y . Yun, H. Lee, G. Bang, and Y . Uh, “Per-gaussian embedding-based deformation for deformable 3d gaussian splatting,” in European Conference on Computer Vision (ECCV), 2024

work page 2024
[14]

Optimized 3d gaussian splatting using coarse-to-fine image frequency modulation,

U. Farooq, J.-Y . Guillemaut, G. Thomas, A. Hilton, and M. V olino, “Optimized 3d gaussian splatting using coarse-to-fine image frequency modulation,” inProceedings of the 22nd ACM SIGGRAPH European Conference on Visual Media Production, ser. CVMP ’25. New York, NY , USA: Association for Computing Machinery, 2025. [Online]. Available: https://doi.org/10.1...

work page doi:10.1145/3756863.3769707 2025
[15]

From coarse to fine: Learn- able discrete wavelet transforms for efficient 3d gaussian splatting,

H. Nguyen, A. Le, B. R. Li, and T. Nguyen, “From coarse to fine: Learn- able discrete wavelet transforms for efficient 3d gaussian splatting,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, October 2025, pp. 3139–3148

work page 2025
[16]

Strang and T

G. Strang and T. Nguyen,Wavelets and filter banks. SIAM, 1996

work page 1996
[17]

Masked wavelet representation for compact neural radiance fields,

D. Rho, B. Lee, S. Nam, J. C. Lee, J. H. Ko, and E. Park, “Masked wavelet representation for compact neural radiance fields,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 20 680–20 690

work page 2023
[18]

Wavenerf: Wavelet-based generalizable neural radiance fields,

M. Xu, F. Zhan, J. Zhang, Y . Yu, X. Zhang, C. Theobalt, L. Shao, and S. Lu, “Wavenerf: Wavelet-based generalizable neural radiance fields,”

work page
[19]

Available: https://arxiv.org/abs/2308.04826

[Online]. Available: https://arxiv.org/abs/2308.04826

work page arXiv
[20]

Trinerflet: A wavelet based triplane nerf rep- resentation,

R. Khatib and R. Giryes, “Trinerflet: A wavelet based triplane nerf rep- resentation,” 2024. [Online]. Available: https://arxiv.org/abs/2401.06191

work page arXiv 2024
[21]

Dwtnerf: Boosting few-shot neural radiance fields via discrete wavelet transform,

H. Nguyen, B. R. Li, and T. Nguyen, “Dwtnerf: Boosting few-shot neural radiance fields via discrete wavelet transform,” 2025. [Online]. Available: https://arxiv.org/abs/2501.12637

work page arXiv 2025
[22]

Instant neural graphics primitives with a multiresolution hash encoding,

T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM Trans. Graph., vol. 41, no. 4, pp. 102:1–102:15, Jul. 2022. [Online]. Available: https://doi.org/10.1145/3528223.3530127

work page doi:10.1145/3528223.3530127 2022
[23]

Micro-macro wavelet-based gaussian splatting for 3d reconstruction from unconstrained images,

Y . Li, C. Lv, H. Yang, and D. Huang, “Micro-macro wavelet-based gaussian splatting for 3d reconstruction from unconstrained images,”

work page
[24]

Available: https://arxiv.org/abs/2501.14231

[Online]. Available: https://arxiv.org/abs/2501.14231

work page arXiv
[25]

Wavelet- gs: 3d gaussian splatting with wavelet decomposition,

B. Zhao, Y . Zhou, S. Yu, Z. Wang, and H. Wang, “Wavelet- gs: 3d gaussian splatting with wavelet decomposition,” in Proceedings of the 33rd ACM International Conference on Multimedia, ser. MM ’25. New York, NY , USA: Association for Computing Machinery, 2025, p. 8616–8625. [Online]. Available: https://doi.org/10.1145/3746027.3755589

work page doi:10.1145/3746027.3755589 2025
[26]

Dwtgs: Rethinking frequency regularization for sparse-view 3d gaussian splatting,

H. Nguyen, R. Li, A. Le, and T. Nguyen, “Dwtgs: Rethinking frequency regularization for sparse-view 3d gaussian splatting,” inProceedings of the 2025 IEEE International Conference on Visual Communications and Image Processing (VCIP). IEEE, 2025

work page 2025
[27]

3d-gsw: 3d gaussian splatting watermark for protecting copyrights in radiance fields,

Y . Jang, H. Park, F. Yang, H. Ko, E. Choo, and S. Kim, “3d-gsw: 3d gaussian splatting watermark for protecting copyrights in radiance fields,”arXiv preprint arXiv:2409.13222, 2024

work page arXiv 2024
[28]

Waveletgaussian: Wavelet- domain diffusion for sparse-view 3d gaussian object reconstruction,

H. Nguyen, R. Li, A. Le, and T. Nguyen, “Waveletgaussian: Wavelet- domain diffusion for sparse-view 3d gaussian object reconstruction,” in Proceedings of the 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026

work page 2026
[29]

Efficient multi-scale network with learnable discrete wavelet transform for blind motion deblurring,

X. Gao, T. Qiu, X. Zhang, H. Bai, K. Liu, X. Huang, H. Wei, G. Zhang, and H. Liu, “Efficient multi-scale network with learnable discrete wavelet transform for blind motion deblurring,” 2024. [Online]. Available: https://arxiv.org/abs/2401.00027

work page arXiv 2024
[30]

A novel learnable orthogonal wavelet unit neural network with perfection reconstruction constraint relaxation for image classification,

A. D. Le, S. Jin, Y . S. Bae, and T. Nguyen, “A novel learnable orthogonal wavelet unit neural network with perfection reconstruction constraint relaxation for image classification,” in2023 IEEE International Confer- ence on Visual Communications and Image Processing (VCIP), 2023, pp. 1–5

work page 2023
[31]

A lattice-structure-based trainable orthogonal wavelet unit for image classification,

A. D. Le, S. Jin, Y .-S. Bae, and T. Q. Nguyen, “A lattice-structure-based trainable orthogonal wavelet unit for image classification,”IEEE Access, vol. 12, pp. 88 715–88 727, 2024

work page 2024
[32]

Biorthogonal lattice tunable wavelet units and their implementation in convolutional neural networks for computer vision problems,

A. D. Le, S. Jin, S. Seo, Y .-S. Bae, and T. Q. Nguyen, “Biorthogonal lattice tunable wavelet units and their implementation in convolutional neural networks for computer vision problems,”IEEE Open Journal of Signal Processing, pp. 1–16, 2025

work page 2025
[33]

Biorthogonal tun- able wavelet unit with lifting scheme in convolutional neural network,

A. Le, H. Nguyen, S. Seo, Y .-S. Bae, and T. Nguyen, “Biorthogonal tun- able wavelet unit with lifting scheme in convolutional neural network,” inProceedings of the 33rd European Signal Processing Conference (EUSIPCO), 2025

work page 2025
[34]

Stop-band energy constraint for orthogonal tunable wavelet units in convolutional neural networks for computer vision problems,

A. D. Le, H. Nguyen, S. Seo, Y .-S. Bae, and T. Q. Nguyen, “Stop-band energy constraint for orthogonal tunable wavelet units in convolutional neural networks for computer vision problems,” inProceedings of the 2025 IEEE International Conference on Visual Communications and Image Processing (VCIP). IEEE, 2025

work page 2025
[35]

Universal wavelet units in 3d retinal layer segmentation,

A. D. Le, H. Nguyen, M. Tran, J. Most, D.-U. G. Bartsch, W. R. Freeman, S. Borooah, T. Q. Nguyen, and C. An, “Universal wavelet units in 3d retinal layer segmentation,” 2025. [Online]. Available: https://arxiv.org/abs/2507.16119

work page arXiv 2025
[36]

Tunable wavelet unit based convolutional neural network in optical coherence tomography analysis enhancement for classifying type of epiretinal membrane surgery,

A. Le, N. Mehta, W. Freeman, I. Nagel, M. Tran, A. Heinke, A. Ag- nihotri, L. Cheng, D.-U. Bartsch, H. Nguyen, T. Nguyen, and C. An, “Tunable wavelet unit based convolutional neural network in optical coherence tomography analysis enhancement for classifying type of epiretinal membrane surgery,” inProceedings of the 33rd European Signal Processing Confere...

work page 2025
[37]

Local light field fusion: Practical view synthesis with prescriptive sampling guidelines,

B. Mildenhall, P. P. Srinivasan, R. Ortiz-Cayon, N. K. Kalantari, R. Ra- mamoorthi, R. Ng, and A. Kar, “Local light field fusion: Practical view synthesis with prescriptive sampling guidelines,”ACM Transactions on Graphics (TOG), 2019

work page 2019
[38]

Mip-nerf 360: Unbounded anti-aliased neural radiance fields,

J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Mip-nerf 360: Unbounded anti-aliased neural radiance fields,”CVPR, 2022

work page 2022
[39]

The unreasonable effectiveness of deep features as a perceptual metric,

R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in CVPR, 2018

work page 2018
[40]

Wavecnet: Wavelet integrated cnns to suppress aliasing effect for noise-robust image classification,

Q. Li, L. Shen, S. Guo, and Z. Lai, “Wavecnet: Wavelet integrated cnns to suppress aliasing effect for noise-robust image classification,”IEEE Transactions on Image Processing, vol. 30, pp. 7074–7089, 2021

work page 2021