Writing study: Selfi: Self-improving Reconstruction Engine via 3D Geometric Feature Alignment

CVPR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← CVPR 2026 oral writing corpus

Youming Deng, Songyou Peng, Junyi Zhang, Kathryn Heal, Tiancheng Sun, John Flynn, Steve Marschner, Lucy Chai · 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: Training, objectives & optimization (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

11 physical pages

Before references / appendix

≈ 8.11 page-equivalents

Main figures / tables

6 / 6

Attached appendix figures / tables

0 / 0

Attached appendix span

Separate file recorded below; not in this PDF total

Before references / appendix: ≈ 73.73%References / intervening material: ≈ 26.27%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	1	≈ 0.63 pp.	5.73%
2. Related Work	2	≈ 0.85 pp.	7.73%
3. Method	3	≈ 2.08 pp.	18.91%
4. Experiments	5	≈ 3.49 pp.	31.73%
5. Conclusion	8	≈ 0.3 pp.	2.73%
References	9	≈ 2.89 pp.	26.27%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 3
• Figure 3 · p. 5
• Figure 4 · p. 5
• Table 1 · p. 6
• Figure 5 · p. 6
• Table 2 · p. 7
• Table 3 · p. 7
• Table 4 · p. 7
• Table 5 · p. 8
• Figure 6 · p. 8
• Table 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 8 physical pages; 7 detected figures and 3 tables. These may include overlapping main-text material.

Located appendix only: starts on physical p.1; ≈ 7.93 page-equivalents, 7 figure labels and 3 table labels. This interval is distinct from the whole file.

Detected heading	PDF page	Estimated span
Supplementary Material	1	≈ 0.07 pp.
1. Supplementary Methods	1	≈ 0.76 pp.
2. Supplementary Experiments	1	≈ 3.21 pp.
3. Limitations	5	≈ 0.42 pp.
References	5	≈ 3.45 pp.

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

Argument and organization

1. Problem and contribution

   Selfi distinguishes foundation-model geometry predictions from geometrically aligned feature spaces. The proposed alignment is motivated as useful both for rendering and for improving estimated cameras, linking the objective to two downstream roles.

2. Argument sequence

   Section 3 explicitly previews its three steps: geometric feature alignment, Gaussian prediction and dense bundle adjustment. Experiments separately vary sequence length, overlap and refinement, following that causal decomposition.

3. Evidence and verification

   Table 1 groups 6/12/24/36-frame inputs, while separate experiments vary overlap. The caption explicitly marks per-scene 3DGS with ground-truth cameras as an upper bound, avoiding a false assumption that all rows share inputs and optimization.

Problem, claim and contribution

Selfi distinguishes foundation-model geometry predictions from geometrically aligned feature spaces. The proposed alignment is motivated as useful both for rendering and for improving estimated cameras, linking the objective to two downstream roles.

Introduction; Section 3.1 · Main paper · Additional source 2

Argument and section order

Section 3 explicitly previews its three steps: geometric feature alignment, Gaussian prediction and dense bundle adjustment. Experiments separately vary sequence length, overlap and refinement, following that causal decomposition.

Section 3; Sections 4.2–4.3 · Main paper · Additional source 2

Figures, captions and visual explanation

Figure 2 identifies a frozen VGGT source of pseudo-depth/cameras and a trainable adapter, then draws the reprojection correspondence underlying the loss. The diagram makes clear that the supervisory geometry is self-generated, not an external ground-truth sensor.

Figure 2, PDF p.3 · Main paper · Additional source 2

Tables, comparisons and evidence

Table 1 groups 6/12/24/36-frame inputs, while separate experiments vary overlap. The caption explicitly marks per-scene 3DGS with ground-truth cameras as an upper bound, avoiding a false assumption that all rows share inputs and optimization.

Table 1, PDF p.6; Table 2 · Main paper · Additional source 2

Appendix and reproducibility

The supplement records alternate objectives, including a contrastive alternative, and expands feature processing, evaluation protocols, bundle-adjustment settings and limitations. Reporting an unsuccessful alternative helps explain the adopted loss rather than only documenting the winner.

Supplementary objective comparison and implementation/limitations sections, PDF pp.1–8 · Main paper · Additional source 2

Clarity and writing conventions

The writing labels pseudo-ground-truth repeatedly and explains which module is frozen versus optimized. This is important precision for a method whose name suggests self-improvement but whose initial supervision can still inherit backbone errors.

Figure 2 caption; Section 3.1 · Main paper · Additional source 2

What to reuse—and what not to infer

Transferable lesson: Show exactly where self-supervision comes from, mark frozen/trainable components, and vary view count separately from overlap when testing multiview reasoning.

Boundary: Agreement with self-generated correspondences is not independent evidence of geometric truth; keep camera access, per-scene optimization and foundation-model dependencies explicit.

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.9–11: abstract/framing, section sequence, figure/table captions, evaluation organization and conclusions; not an exhaustive technical verification.

Appendix / supplement

Writing-oriented supplement scan: all eight supplementary PDF pages, including alternate objectives, protocols, geometry refinement, limitations and extra qualitative results. Video/code assets were not executed or assessed.

Visually inspected PDF pages

3, 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 · 11 pages
• Official CVF supplementary PDF · 8 pages