Writing study: Proxy-GS: Unified Occlusion Priors for Training and Inference in Structured 3D Gaussian Splatting

CVPR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← CVPR 2026 oral writing corpus

Yuanyuan Gao, YUNING GONG, Yifei Liu, Jingfeng Li, Dan Xu, Yanci Zhang, Dingwen Zhang, Xiao Sun, Zhihang Zhong · Official CVPR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Efficiency, inference & systems (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

10 physical pages

Before references / appendix

≈ 8.06 page-equivalents

Main figures / tables

6 / 4

Attached appendix figures / tables

0 / 0

Attached appendix span

Separate file recorded below; not in this PDF total

Before references / appendix: ≈ 80.6%References / intervening material: ≈ 19.4%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1. Introduction	2	≈ 0.65 pp.	6.5%
2. Related Work	2	≈ 0.86 pp.	8.6%
3. Preliminaries	3	≈ 0.5 pp.	5%
4. Method	4	≈ 1.76 pp.	17.6%
5. Experiment	5	≈ 2.95 pp.	29.5%
6. Conclusion	8	≈ 0.12 pp.	1.2%
7. Acknowledgment	8	≈ 0.14 pp.	1.4%
References	9	≈ 1.94 pp.	19.4%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 5
• Figure 3 · p. 5
• Table 1 · p. 6
• Table 2 · p. 6
• Figure 4 · p. 7
• Table 3 · p. 7
• Figure 5 · p. 8
• Table 4 · p. 8
• Figure 6 · p. 8

Attached appendix

No recognized caption labels.

Separate source files—excluded from canonical-PDF totals

A separate author version may repeat main text and references. Whole-file lengths and counts are not extra appendix material and are never added to the canonical PDF totals. Read the study's source scope for which sections received substantive reading versus layout inspection.

Additional source 1

Separate PDF (or ZIP containing it): whole-file layout of 4 physical pages; 3 detected figures and 6 tables. These may include overlapping main-text material.

Located appendix only: starts on physical p.1; ≈ 3.85 page-equivalents, 3 figure labels and 6 table labels. This interval is distinct from the whole file.

Detected heading	PDF page	Estimated span
Supplementary Material	1	≈ 0.3 pp.
1. Appendix	1	≈ 2.25 pp.
2. Cluster construction. We partition theFsimplified mesh	3	≈ 1.31 pp.

• Table 1 · p. 1
• Table 2 · p. 1
• Figure 2 · p. 1
• Table 3 · p. 2
• Table 4 · p. 2
• Table 5 · p. 3
• Figure 3 · p. 3
• Table 6 · p. 3
• Figure 4 · p. 4

Argument and organization

1. Problem and contribution

   Proxy-GS argues that processing occluded anchors wastes work and that the same geometric prior can improve training placement. The contribution is framed around a common bottleneck across training and rendering, not only faster rasterization.

2. Argument sequence

   Background on structured Gaussian decoding and hardware rasterization precedes the proxy construction, filtering and densification interventions. Experiments then move from scene-level quality/speed to switches and proxy-fidelity sensitivity.

3. Evidence and verification

   Tables 1–2 pair PSNR/SSIM/LPIPS with FPS and divide scenes by setting. The text specifies A100 training and RTX4090 inference, while the component table independently toggles training/inference uses of the prior.

Problem, claim and contribution

Proxy-GS argues that processing occluded anchors wastes work and that the same geometric prior can improve training placement. The contribution is framed around a common bottleneck across training and rendering, not only faster rasterization.

Introduction; Sections 3–4 · Main paper · Additional source 2

Argument and section order

Background on structured Gaussian decoding and hardware rasterization precedes the proxy construction, filtering and densification interventions. Experiments then move from scene-level quality/speed to switches and proxy-fidelity sensitivity.

Sections 3–5 · Main paper · Additional source 2

Figures, captions and visual explanation

Figure 2 visually separates preprocessing, hardware proxy-depth generation, anchor densification and culling/decoding. Figure 3 breaks total inference time into rendering, filtering and depth components, showing where the speed gain comes from.

Figures 2–3, PDF p.5 · Main paper · Additional source 2

Tables, comparisons and evidence

Tables 1–2 pair PSNR/SSIM/LPIPS with FPS and divide scenes by setting. The text specifies A100 training and RTX4090 inference, while the component table independently toggles training/inference uses of the prior.

Tables 1–3, PDF pp.6–7; implementation paragraph · Main paper · Additional source 2

Appendix and reproducibility

The supplement gives MatrixCity partitions, proxy depth acquisition, decoded-anchor and noise analyses, mesh extraction details and hardware-rendering combinations. These details make proxy quality and integration overhead auditable.

Supplementary Sections 1.1–1.7, PDF pp.1–4 · Main paper · Additional source 2

Clarity and writing conventions

The writing states where occlusion is strong or weak, instead of implying that one maximum speedup applies everywhere. The mesh-resolution and noise analyses connect that deployment condition to observed trade-offs.

Tables 1–2; Figures 5–6, main PDF p.8 · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: For an accelerator, show an end-to-end time decomposition and plot quality against the accuracy of the auxiliary representation it depends on.

Boundary: Speedup varies with scene occlusion, proxy quality and hardware. A faster anchor stage alone is not the same as the same multiplier for total rendering.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-26

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan throughout main PDF pp.1–8 and references pp.9–10: abstract/framing, section sequence, figure/table captions, evaluation organization and conclusions; not an exhaustive technical verification.

Appendix / supplement

Writing-oriented supplement scan: all four supplementary PDF pages covering dataset blocks, hardware integration, proxy depth/mesh details and sensitivity analyses. Video/code assets were not executed or assessed.

Visually inspected PDF pages

5, 6

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Official CVF open-access CVPR 2026 accepted-paper PDF · 10 pages
• Official CVF supplementary PDF · 4 pages