brycewang-stanford ECCV Writing Style
Use this while drafting or compressing an ECCV manuscript. The defining constraint is the format: LNCS single-column, 14 pages including figures and tables, references-only pages after that. This is a different craft from the two-column CVF page — more words per page, cheaper full-width figures, costlier white space — and the template changed after 2024, so write in the current kit from day one.
Spending 14 single-column pages
| Budget line | Typical allocation | LNCS-specific note |
|---|---|---|
| Title, abstract, intro | 2.0 pages | Single column makes long intros tempting; cap at 1.25 pages of prose |
| Related work | 1.0 page | Numbered [n] citations pack tightly; group by idea, not by paper |
| Method | 3.5–4.5 pages | Full-width architecture figures are cheap here — use one, not three |
| Experiments | 4.5–5.5 pages | Tables span the full text width; design them wide and few |
| Limitations + conclusion | 0.5–1.0 page | An honest limitations paragraph is expected |
| Figures/tables total | counted inside the 14 | The submission limit includes them — every figure displaces prose |
The inclusion of figures in the limit is the discipline lever: each figure must either carry an argument (teaser, method overview, failure modes) or compress one (results curves). Decorative variants go to the supplement.
First-page architecture
An ECCV first page must work for the adjacent-niche reviewer (see
eccv-review-process): by the end of page 1 a non-specialist should be able
to say what input becomes what output, why current methods fail at it, and
what single idea fixes it.
- Teaser figure: input → output with one baseline failure alongside, captioned so it is understandable without the body text.
- Contribution list: two or three bullets, each falsifiable ("we show X improves Y by Z on W"), none of them "extensive experiments".
- Name the setting precisely (supervised? zero-shot? per-scene?) — mixed panels misfile vague papers into the wrong baseline universe.
LNCS mechanics that bite
- Springer's numbered-citation style makes "[23] showed" unreadable; write "Author et al. [23] showed" so prose survives the bibliography.
\subsubsectiondepth is where LNCS headings stop looking like headings; restructure rather than nest.- Single-column tables tolerate more columns than two-column CVF tables — merge your ablation grids instead of stacking small tables.
- Keep the references-only rule literal: nothing but bibliography beyond the limit — a stray appendix line on page 15 is a compliance failure.
Claim phrasing sized for the rebuttal
Every headline sentence should be defensible in two rebuttal lines later:
Fragile: "Our method significantly outperforms all prior work."
Sized: "On <benchmark vX>, our method improves <metric> by <d> over
<strongest baseline> under matched <backbone/training data>."
Fragile: "generalizes to arbitrary scenes"
Sized: "generalizes across the N categories of <dataset>; unseen
domains are evaluated in Sec. 5.4 and remain open"
Scope every superlative to benchmark + protocol + substrate. In a foundation-model era, unmatched pretraining data is the first rebuttal attack surface — declare the substrate in the claim itself.
Revision passes, in order
- Argument pass: does each section answer the previous one's open question?
- Figure pass: captions self-contained; every figure referenced; each earns its page share against the 14-page count.
- Claim pass: apply the sizing patterns above to abstract, intro, and conclusion — the three places reviewers quote from.
- Anonymity pass: self-citations in third person, no repo links, no acknowledgements.
- Template pass: current-cycle kit, margins untouched, references-only overflow.
Output format
[Style verdict] submit-ready / restructure / compress / re-scope claims
[Page ledger] <section -> pages vs budget, figure share>
[First-page test] <what an adjacent-niche reader retains, or fails to>
[Fragile claims] <sentence -> sized rewrite>
[Pass order] <which of the five passes remain>