brycewang-stanford ECCV Supplementary
Use this when planning the ECCV supplement. ECCV 2026 kept the classic vision-conference structure the CVF siblings have moved away from: the supplement had its own deadline a week after the paper (paper March 5, supplement March 12). That week is a gift with sharp edges — the paper is already frozen, so the supplement can only support claims, never add them.
What the extra week is actually for
Legitimate uses of March 5 → March 12:
- Rendering the qualitative-results PDF and video reels at full quality.
- Finishing the code archive scrub (
eccv-artifact-evaluation). - Formatting extended tables and per-category breakdowns already summarized in the body.
Illegitimate uses that reviewers punish:
- Landing a result the body cites as "see supplement" but never summarizes — the 14 pages must stand alone.
- Squeezing overflow method description out of the body; a reader of only the main PDF must be able to reimplement the core.
- Adding experiments the frozen body never mentions, hoping reviewers credit them anyway.
Supplement inventory for a vision paper
| Component | Why it earns its place | Packaging rule |
|---|---|---|
| Qualitative grids (success + failure) | Vision reviewers trust pixels over deltas | Same scenes across all baselines; no cherry-pick without a random-sample page |
| Video results | Temporal methods are unjudgeable from stills | Short, compressed, playable without exotic codecs |
| Per-dataset / per-category tables | Backs the body's averaged claims | Every table cross-referenced from the body |
| Full training recipes + configs | Reproducibility ledger overflow | Mirror the code archive; no contradictions with body numbers |
| Extra ablations | Answers the predictable "what about X?" review | One line each in the body's ablation discussion |
| Proof/derivation details | Rare at ECCV but present in geometry papers | Notation identical to the body |
Anonymity scrub for supplement archives
Supplement uploads leak identity through channels the paper PDF does not:
□ Video files: no institutional watermarks, no title slides with names
□ MP4/PDF metadata: author, tool, and GPS fields wiped
□ Code: paths (/home/<user>/, /cluster/<lab>/), git config, W&B entities
□ Screenshots: browser profiles, OS usernames, bookmark bars cropped out
□ Filenames: no lab acronyms or grant codes in archive member names
□ Prior-work assets: reused figures that only your group could possess
Reviewer-attention model
Plan around discretionary reading: the qualitative PDF gets skimmed by most reviewers, videos get opened when the paper is temporal, tables get opened when a body claim looks suspicious, and code gets opened by at most one reviewer. Therefore: put the persuasion (visual evidence) where skimming lands, and the protection (recipes, breakdowns) where suspicion lands.
Size and format constraints
Upload caps and allowed formats are set per cycle in OpenReview and were not re-verified for 2026 (待核实) — check the submission form early, because video compression to meet a cap takes an evening you do not want to discover on March 11. Keep one master script that rebuilds the whole supplement archive from sources so a size cut does not require manual re-assembly.
Output format
[Supplement verdict] ready / thin / overloaded / body-dependent
[Inventory] <component -> present, cross-referenced from body?>
[Self-containment check] <body claims that lean on supplement-only material>
[Anonymity scrub] <leaks found -> fix>
[Deadline plan] <what lands March 5 vs the supplement week>