Writing study: LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models

ICCV overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICCV 2025 oral writing corpus

Haiwen Huang, Anpei Chen, Volodymyr Havrylov, Andreas Geiger, Dan Zhang · Official ICCV 2025 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

11 physical pages

Before references / appendix

≈ 8.1 page-equivalents

Main figures / tables

8 / 6

Attached appendix figures / tables

0 / 0

Attached appendix span

No boundary detected; not assumed zero

Before references / appendix: ≈ 73.64%References / intervening material: ≈ 26.36%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	≈ 1.57 pp.	14.27%
2. Related Work	2	≈ 0.81 pp.	7.36%
3. Coordinate-Based Feature Representation	3	≈ 0.59 pp.	5.36%
4. Training Objective	4	≈ 2.55 pp.	23.18%
5. Experiments	6	≈ 2 pp.	18.18%
6. Conclusion	8	≈ 0.21 pp.	1.91%
Acknowledgement	9	≈ 0.06 pp.	0.55%
References	9	≈ 2.9 pp.	26.36%

All detected figure and table locations

Main text

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

Attached appendix

No recognized caption labels.

Argument and organization

1. 1 Introduction; 2 Related Work

   Separates upsampler capacity from the lack of high-resolution task-agnostic supervision.

2. 3 Coordinate-Based Feature Representation; 4 Training Objective

   Introduces global cross-attention, then mask-refined pseudo-targets and crop-based self-distillation.

3. 5 Experiments; 6 Conclusion

   Evaluates six task families and explicitly crosses architecture with training-objective choices before scale/runtime checks.

Problem, claim and contribution

Treats high-resolution feature learning as both an architecture problem and a supervision-resolution problem, rather than assuming interpolation alone fixes dense prediction.

Sections 1 and 4, physical pp.1–2,4 · Main paper

Argument and section order

The objective section first explains how prior low-resolution losses underconstrain outputs, then builds two increasingly refined pseudo-target stages.

Sections 4.1–4.2, pp.4–6 · Main paper

Figures, captions and visual explanation

The training diagram separates mask-assisted stage one from teacher/student stage two and explicitly notes that an omitted teacher VFM block is only a drawing simplification.

Figure 4, p.4 · Main paper

Tables, comparisons and evidence

The architecture table fixes the LoftUp objective while the adjacent chart applies the new pseudo-targets to other upsamplers, separating capacity from supervision effects.

Table 4 and Figure 7, p.7 · Main paper

Appendix and reproducibility

Alternative CLIP/RADIO backbones, training details, additional predictions and architecture ablations are assigned to supplements, including Table B.3. Only main pointers were checked.

Sections 5 and 5.2, pp.6–8 · Main paper

Clarity and writing conventions

Input resolution and final feature resolution are explicitly distinguished in the scale study, preventing a larger output grid from being confused with a larger VFM input.

Table 5 and Section 5.2, p.8 · Main paper

What to reuse—and what not to infer

Transferable lesson: When two ingredients improve together, cross the new objective with old architectures and the new architecture with fixed supervision.

Boundary: Pseudo-ground-truth features are not independent true dense features, and SAM masks contribute a learned structural prior. Task-agnostic training does not establish equal gains on every possible task; A100 upsampler latency is not the entire backbone-plus-head pipeline.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented scan across the main manuscript, physical PDF pp.1–8, covering framing, section progression, method explanations, captions, comparison protocols and concluding scope; reference boundary checked. Selected full-page rasters were inspected. Not exhaustive line-by-line reading, proof verification or experimental reproduction.

Appendix / supplement

Main-paper appendix/supplement references were checked. A separate supplementary source is linked in the inventory but was not retrieved or inspected; its contents are not included in the reading or visual-page counts.

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.

• Official CVF open-access ICCV 2025 accepted-paper PDF, retrieved 7 October 2026; CVF states identical to accepted version except watermark; not assumed identical to initial submission · 11 pages