PaperlayerView exact revision on arXiv

Source-linked AI summary

ADATEX4D: adaptive texture capacity allocation for 4D gaussian splatting

De Jiang, Peiqiang Wang, Kehong Yuan, Shaohua Ma

arXiv:2609.29963v1cs.CVcs.AI

TL;DR

Uniform textures spend the same capacity on dynamic Gaussians despite differences in visibility and local detail. AdaTex4D allocates packed textures anisotropically using visibility-aware temporal evidence and deformed scales while freezing the 4DGS backbone. Across the reported scenes it uses less than half the uniform texture bytes while preserving reconstruction quality, and fixed-memory comparisons report quality gains over uniform assignment.

  • Problem

    Uniform texture assignment gives every primitive the same square texture despite variation in projected footprint, visibility, motion, and local detail, leaving the allocation of texture capacity unresolved.

    Source: §1 · p. 1

  • Method

    AdaTex4D freezes a trained deformation-based 4DGS backbone and optimizes packed RGBA residual textures grown anisotropically from visibility-normalized gradients, temporal peaks, and deformed local scales.

    Source: §1 · p. 1

  • Results

    47.3% of uniform texture bytes preserves reconstruction quality across 12 scenes, while at 80 MiB AdaTex4D improves PSNR by 0.175 and 0.205 dB over Uniform Textured 4DGS on N3DV and PanopticSports, respectively.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Fixed-memory and fixed-Gaussian-count protocols separate the benefit of adaptive placement from simply increasing texture resolution or primitive count.

    Source: §1 · p. 1

    The current renderer materializes dense textures, limiting throughput gains from packed storage.

    Source: §2.4 · p. 2

Abstract

from arXiv · show

Textured Gaussians improve local appearance capacity, but assigning the same texture resolution to every primitive wastes storage on low-detail or weakly visible regions. We introduce AdaTex4D, an adaptive texture-capacity module for deformation-based 4D Gaussian Splatting. Each Gaussian carries packed RGBA triplanes whose two axes grow independently according to visibility normalized screen-space gradients and deformed local scales. Experiments on N3DV and PanopticSports show that AdaTex4D reduces texture storage by more than half while preserving reconstruction quality. Under fixed memory budgets, adaptive allocation also improves quality over uniform texture assignment and reduces overall model and peak memory. These results show that dynamic, anisotropic texture allocation provides a more efficient way to distribute local appearance capacity in 4D Gaussian representations.

1. INTRODUCTION

Uniform textures spend the same capacity on dynamic Gaussians despite differences in visibility, motion, and local detail. AdaTex4D uses visibility-aware temporal evidence and deformed scales to allocate anisotropic texture capacity while freezing the 4DGS backbone; experiments report lower texture storage and improved quality under fixed memory budgets.

1. INTRODUCTION

  • Uniform textures charge small, occluded, or visually simple Gaussians the same as detail-critical regions, while increasing global resolution raises memory cost for every primitive.The allocation question is where to place texture capacity.Source: §1 · p. 1
  • Dynamic allocation must account for visibility, changing deformed scales, and short-lived detail peaks without reacting to transient per-frame noise.Canonical geometry misses time-varying states, while simple temporal averages can suppress rare peaks.Source: §1 · p. 1
  • AdaTex4D freezes geometry, deformation, opacity, SH appearance, and primitive count, and grows packed RGBA triplanes anisotropically using visibility-normalized gradients, temporal peaks, and deformed local scales.The three planes start at 1 × 1 and grow to at most 4×4.Source: §1 · p. 1, Figure 2 · p. 2
  • The reported storage and memory measures reflect different costs, and dense texture materialization in the current renderer limits throughput gains from packing.The paper reports checkpoint size, peak GPU memory, and runtime alongside packed texel counts.Source: §1 · p. 1
  • At 80 MiB, AdaTex4D improves PSNR by 0.175 dB on N3DV and 0.205 dB on PanopticSports over Uniform Textured 4DGS; across 12 scenes, it uses 47.3% of uniform texture bytes while preserving reconstruction quality.Source: §1 · p. 2

2. METHOD

AdaTex4D adds packed RGBA texture planes to a frozen 4DGS backbone and allocates capacity using visibility-aware temporal gradient demand and deformed scales. Anisotropic growth is bounded, and optimization and storage measurements account for the renderer’s dense texture materialization.

2.1. Dynamic Gaussian parameterization

  • A 3D sample is mapped into the Gaussian’s local frame, projected onto three RGBA planes, and sampled to form a residual appearance feature while alpha modulates opacity.The geometry, deformation field, and Gaussian count are frozen; only the texture branch is optimized.Source: §2.1 · p. 2

2.2. Visibility-normalized dynamic evidence

  • Visible occurrences use projected-center gradient magnitude as texture demand, aggregated through means over eight temporal bins into persistent and peak demand.Only visible observations contribute; the highest-demand bin also supplies the mean deformed scale for growth.Source: §2.2 · p. 2

2.3. Anisotropic texture growth

  • At growth steps 500 and 1000, a Gaussian resizes only when its allocation score exceeds 2 × 10−5, doubling one or both axes according to selected-bin deformed-scale ratios.Each doubled axis is capped at 4; new texels and Adam moments are resized by area interpolation.Source: §2.3 · p. 2

2.4. Optimization and storage

  • Packed storage scales with active RGBA texels plus dimension and offset metadata, but the renderer materializes dense 4×4 textures, so packing does not yet increase rendering throughput.The study also reports full-model storage, peak allocated memory, training time, and FPS.Source: §2.4 · p. 2
  • Figure 3 compares methods by column, with full frames above enlarged regions of interest for Tennis and Sear Steak.The figure’s top scene is Tennis and its bottom scene is Sear Steak.Source: Figure 3 · p. 3

3. EXPERIMENTS

Across 12 N3DV and PanopticSports scenes, AdaTex4D preserves reconstruction quality with less texture, full-model, and peak memory than Uniform Textured 4DGS. Fixed-budget and fixed-count comparisons, qualitative examples, and ablations examine where the gains hold and which allocation signals matter.

3.1. Protocol

  • The study evaluates six N3DV and six PanopticSports scenes, comparing 4DGaussians, Uniform Textured 4DGS, and AdaTex4D under fixed-memory and fixed-count protocols.Fixed-memory budgets are 80 and 150 MiB; fixed-count comparisons use 25% and 50% of the converged 4DGaussians count with identical Gaussian subsets across methods.Source: §3.1 · p. 3

3.2. Main quality–storage result

  • 52.7% less texture storage accompanies similar quality across 12 scenes: AdaTex4D uses 51.7 versus 109.3 MiB, with 24.4% less full-model storage and 27.0% less peak memory.PSNR is 28.191/28.157 dB (+0.034 dB; paired 95% bootstrap CI [−0.035, 0.113]); training takes 2.18/1.64 h and rendering is 84.1/100.8 FPS for AdaTex4D/Uniform Textured 4DGS.Source: §3.2 · p. 3

3.3. Comparison under fixed memory budgets

  • At 80 MiB, AdaTex4D exceeds Uniform Textured 4DGS by 0.175/0.205 dB on N3DV/PanopticSports; at 150 MiB, PanopticSports 4DGaussians leads PSNR/SSIM while AdaTex4D leads LPIPS.At 80 MiB, AdaTex4D also exceeds 4DGaussians by 0.153/0.152 dB; the reported 150 MiB AdaTex4D PSNR is 27.203 dB on N3DV and 26.532 dB on PanopticSports.Source: §3.3 · p. 3, Table 1 · p. 4

3.4. Comparison at fixed Gaussian counts

  • At 25%/50% Gaussian count, AdaTex4D saves 16.9%/17.3% active-model storage on N3DV and 19.4%/19.2% on PanopticSports versus Uniform Textured 4DGS.Its PSNR gaps are 1.093/0.683 dB on N3DV and 0.150/0.152 dB on PanopticSports; results indicate texture allocation cannot fully compensate for missing Gaussian support at 25%.Source: Table 2 · p. 4, §3.4 · p. 4

3.5. Qualitative comparison

  • In Tennis and Sear Steak, AdaTex4D retains head and racket-hoop outlines and steak boundaries and surface detail, unlike the blurrier 4DGaussians and Spacetime Gaussian Feature Splatting results.The retained features resemble Uniform Textured 4DGS despite AdaTex4D using less texture storage.Source: §3.5 · p. 4

3.6. Ablations and allocation behavior

  • Removing visibility normalization lowers PSNR by 0.645/0.652 dB on N3DV/PanopticSports, more than removing temporal peaks or anisotropic allocation.Removing peaks lowers PSNR by 0.331/0.322 dB; removing anisotropic allocation lowers it by 0.461/0.452 dB. Square growth uses 32.4 rather than 30.9 MiB on N3DV and lowers PSNR; 53.8% of Gaussians use an anisotropic plane, and AdaTex4D stores 47.3% of uniform packed texture bytes.Source: §3.6 · p. 4, Table 3 · p. 4

4. CONCLUSION

AdaTex4D allocates anisotropic texture capacity to dynamic Gaussians while freezing the 4DGS backbone. Experiments show comparable reconstruction quality with less storage, though dense texture materialization limits runtime gains.

4. CONCLUSION

  • AdaTex4D achieves comparable reconstruction quality with less storage than uniform texture allocation.Ablations support the allocation signals.Source: §4 · p. 4
  • AdaTex4D grows packed textures anisotropically using visibility-normalized gradients, temporal peaks, and deformed scales while freezing the 4DGS backbone.Source: §4 · p. 4
  • Dense texture materialization currently limits runtime gains.Source: §4 · p. 4

Related papers

Loading 2609.29963v1…