An ICCV submission is read by reviewers drawn from the most international pool in computer vision — in 2025, every author of 11,239 submissions was also a potential reviewer — and most of them work outside your subfield. The style target is therefore precise: a paper a generalist vision researcher can evaluate fairly in one sitting, in their second or third language, under deadline pressure of their own.
The three reviewers you draw will rarely all be specialists in your niche. Concretely:
The cap is eight pages including figures and tables, references-only overflow (2025 rule). Treat layout as a market where every element bids for space with the evidence it delivers:
| Element | Earns its space when | Evict when |
|---|---|---|
| Page-1 overview figure | A stranger states the idea after 30 seconds with it | It needs the body text to decode |
| Pipeline figure | It replaces ≥ its own height in method prose | It restates the obvious encoder→decoder |
| Main results table | Rows = strongest current baselines, matched | Padded with weak ancient rows for bulk |
| Qualitative strip | Shows the mechanism's visible signature | Generic pretty outputs any method makes |
| Ablation table | Each row flips one switch (iccv-experiments) |
Grid dumped without a takeaway sentence |
| Equation block | Notation reused ≥ twice downstream | Formalizing what one sentence says |
Every caption must be self-sufficient — dataset, setting, and the takeaway in words — because the skim pass reads only captions, and at most venues in this family the skim pass happens before the real one.
The genre's strongest habit: name the mechanism, then the consequence, then the number. Compare the two registers:
weak: "We propose a novel and effective framework that significantly
outperforms state-of-the-art methods on several benchmarks."
strong: "Replacing per-frame matching with a persistent memory of object
queries removes the re-detection bottleneck: +3.1 HOTA on
DanceTrack at identical backbone and training budget."
The weak sentence contains zero checkable content — and at ICCV, unfalsifiable praise adjectives ("novel", "significantly", "state-of-the-art" unscoped) are read as inexperience markers. The strong sentence makes three commitments a reviewer can verify, which is exactly what earns scores.
Every claim in the abstract will potentially need defending in May inside a
shared, single rebuttal page (iccv-author-response). Size claims accordingly
at writing time:
Reviewers at volume venues read out of order: abstract → figures → tables → conclusion → (maybe) everything else. Two consequences. First, the paper's argument must be reconstructible from abstract + captions + table headers alone — test this literally by extracting them:
# Skim-pass simulator: what a reviewer sees in the first five minutes
pdftotext -layout paper.pdf - | grep -A2 -E '^(Figure|Table) [0-9]+[.:]' > skim.txt
head -60 skim.txt # does this sequence state problem → idea → evidence?
Second, transitions between sections carry less weight than section openings: begin each section with its conclusion ("Our decoder attends over a persistent memory; this section details the write and read operations"), not with throat- clearing history.
The ICCV intro that works is a proof sketch, not a literature tour: (1) the visual capability at stake, one paragraph, concrete; (2) why current approaches miss it — mechanism of failure, not adjectives; (3) the idea, stated so a generalist could re-derive the design; (4) evidence preview with scoped numbers; (5) contributions as claims (each one falsifiable), not as an activity log ("we conduct extensive experiments" is an activity, not a contribution).
Only after structure settles: strip hedge stacks ("could potentially help to"), convert passive method descriptions to active ("the features are fused" → "the decoder fuses"), unify tense (present for the method, past for experiments run), and hunt duplicated notation — vision papers accumulate three symbols for the same feature map by section 4. If any drafting was machine-assisted, verify every citation exists and every number matches your tables; author responsibility for the text is absolute, and fabricated references are review-killing at every CVF venue.
[Generalist test] adjacent-area reader states the idea after p.1: yes/no
[Space market] elements not earning their area: <list>
[Caption audit] self-sufficient: n/m
[Claim sizing] unscoped superlatives: <locations>
[Skim pass] abstract+captions+headers carry the argument: yes/no
[Edit queue] <highest score-impact first>