Writing study: SAM 3D Animal: Promptable Animal 3D Reconstruction from Images in the Wild

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Xuyi Hu, Jin Lyu, Jiuming Liu, Yebin Liu, Silvia Zuffi, Liang An, Stefan Goetz · Official NeurIPS 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: Architecture & representation (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

16 physical pages

Before references / appendix

≈ 9.09 page-equivalents

Main figures / tables

7 / 3

Attached appendix figures / tables

3 / 1

Attached appendix span

≈ 2.91 page-equivalents

Before references / appendix: ≈ 56.81%References / intervening material: ≈ 25%Attached appendix: ≈ 18.19%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

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	≈ 0.86 pp.	5.38%
2 Related Work	2	≈ 0.88 pp.	5.5%
3 SAM 3D Animal Model	3	≈ 1.78 pp.	11.13%
4 Herd3D Dataset	5	≈ 0.51 pp.	3.19%
5 Experiments	5	≈ 3.94 pp.	24.63%
6 Conclusion	9	≈ 0.34 pp.	2.13%
References	10	≈ 4 pp.	25%
A Bipartite Matching Details	14	≈ 0.53 pp.	3.31%
B More Details about Herd3D	14	≈ 0.17 pp.	1.06%
C Failure Cases in Herd3D Construction.	14	≈ 1.29 pp.	8.06%
D More Experiments	16	≈ 0.91 pp.	5.69%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 4
• Figure 3 · p. 5
• Figure 4 · p. 6
• Figure 5 · p. 7
• Table 1 · p. 7
• Figure 6 · p. 8
• Table 2 · p. 8
• Table 3 · p. 8
• Figure 7 · p. 9

Attached appendix

• Figure 8 · p. 15
• Figure 9 · p. 15
• Table 4 · p. 15
• Figure 10 · p. 16

Argument and organization

1. Problem and contribution

   The introduction turns species diversity, occlusion and multiple interacting animals into explicit requirements for one promptable reconstruction system. It separates full-scene inference from prior pipelines that require individual animal crops.

2. Argument sequence

   SMAL+ preliminaries precede instance-query decoding and matching losses. The new Herd3D dataset then supplies the missing supervision, followed by comparisons, prompt ablations and an occlusion analysis. Data and model changes are distinct explanatory steps.

3. Evidence and verification

   The qualitative gallery pairs inputs with mesh overlays. The adjacent table separates no-prompt, predicted-keypoint and ground-truth-prompt variants, groups columns by dataset and supplies metric directions. These rows prevent an oracle-assisted result from being mistaken for fully automatic performance.

Problem, claim and contribution

The introduction turns species diversity, occlusion and multiple interacting animals into explicit requirements for one promptable reconstruction system. It separates full-scene inference from prior pipelines that require individual animal crops.

Introduction and Sections 3–5, pp. 1–6 · Main paper

Argument and section order

SMAL+ preliminaries precede instance-query decoding and matching losses. The new Herd3D dataset then supplies the missing supervision, followed by comparisons, prompt ablations and an occlusion analysis. Data and model changes are distinct explanatory steps.

Sections 3–5, pp. 3–9 · Main paper

Figures, captions and visual explanation

The architecture figure places the full image, ViT encoder, six-layer query decoder, predicted instances and SMAL+ renderer on one path. Colored matching lines connect predictions to ground-truth instances; token labels make the different geometric and prompt roles inspectable.

Figure 2, p. 4; rendered page inspected · Main paper

Tables, comparisons and evidence

The qualitative gallery pairs inputs with mesh overlays. The adjacent table separates no-prompt, predicted-keypoint and ground-truth-prompt variants, groups columns by dataset and supplies metric directions. These rows prevent an oracle-assisted result from being mistaken for fully automatic performance.

Figure 5 and Table 1, p. 7; rendered page inspected · Main paper

Appendix and reproducibility

The supplement expands bipartite matching and generated-data construction rather than replacing their main-text definitions. It also includes generation failure examples and denser prompting experiments, useful for checking what the synthetic supervision can and cannot establish.

Appendices A–D, pp. 14–16; Figure 9 and Figure 10 · Main paper

Clarity and writing conventions

Named subproblems keep a complex architecture readable, but broad robustness wording should remain tied to tested species, benchmarks and prompt settings. Occlusion plots explicitly define their error bars; their within-group variation should not be relabeled as independent training-seed uncertainty.

Sections 5.2–5.3 and conclusion, pp. 8–9 · Main paper

What to reuse—and what not to infer

Transferable lesson: For promptable models, show the information entering each experimental condition beside the quantitative comparison, and connect architecture components to the task obstacles they address.

Boundary: The inspected source is an author preprint, not a verified accepted revision. Ground-truth prompts and externally predicted prompts supply different information; successful overlays do not establish generalization to every animal family.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-02

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented distributed manuscript scan of Sections 1–6, physical pp. 1–9: section sequence, substantive prose windows, captions, comparison contracts and conclusions. Not exhaustive technical verification.

Appendix / supplement

Writing-oriented distributed scan of Appendices A–D, physical pp. 14–16, including matching, generated-data examples and further experiments. Not independent proof checking; code and experiments were not executed.

Visually inspected PDF pages

4, 7

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.

• Public author arXiv manuscript 2605.07604v1; not verified identical to the accepted or camera-ready OpenReview revision · 16 pages