brycewang-stanford ICCV Supplementary Material
Since the 2025 cycle, ICCV collects the supplement at the paper deadline — one upload event in early March (March 7, 2025, 11:59pm HST; announced as a deliberate process change that also shortened the decision timeline). That single scheduling fact rewrites supplement strategy: there is no packaging week after the paper freezes, so the supplement is a parallel product with its own owner from February, or it is a casualty.
What the same-day rule changes in practice
| Old two-deadline habit | Same-day reality |
|---|---|
| "Ship the paper, then render videos next week" | Videos render while the paper is still being edited — reserve machines for both |
| Supplement absorbs deadline-night leftovers | Anything unfinished at the deadline simply does not exist for reviewers |
| Code scrubbed after the PDF freezes | Identity scan runs on paper and archive the same night |
| Supplement pointers checked at leisure | "See Suppl. C" written on deadline day points at a section that must already exist |
Practical rule: the supplement's content list is frozen when the paper's section structure freezes (about two weeks out), and from then on the archive only executes that list.
Division of labor with the 8 pages
The body must stand alone for a reviewer who never opens the archive; the supplement rewards the reviewer who does. Sort content by the question it answers:
- "Is the claim true?" → body, always: headline tables, the decisive ablation, the honest limitation.
- "Is it true broadly?" → supplement: full ablation grids, additional datasets, per-class or per-sequence breakdowns, sensitivity sweeps.
- "Can I see it?" → supplement, and this is where ICCV differs from text-first venues: qualitative galleries, uncurated sample grids, and video for anything temporal or 3D. State the selection rule in each caption ("validation images 1–12, no curation") — a stated rule is what separates evidence from advertising.
- "Could I rebuild it?" → supplement: configs, pseudo-code, dataset
preprocessing detail, plus the code package (
iccv-artifact-evaluation). - "Does the math hold?" → supplement appendix; vision reviewers spot-check derivations rather than read them linearly, so number every step they might jump to.
Never exile decision-critical content: a reviewer who must open a zip to learn how the method works will instead assume the worst and score accordingly.
Anonymity inside the archive
The archive is reviewed under full double-blind rules, and archives leak differently than PDFs:
# Same-night sweep across the whole supplement
unzip -l supplement.zip | grep -viE '^\s*(Archive|Length|----)' | grep -iE 'user|home|<labname>'
ffprobe -v quiet -show_format videos/main_comparison.mp4 | grep -i 'artist\|author\|comment'
grep -rniE 'wandb|slurm|/nfs/|@(gmail|edu)' code/ --include='*.py' --include='*.yaml' | head
exiftool figures/*.png 2>/dev/null | grep -i 'author\|copyright' | head
Video container tags, W&B entity names in configs, cluster paths in logs, and PNG metadata are the four leaks PDFs never taught you to check. Remember also the 2025 codebase rule: even inside the supplement, do not cite or link your public repository — include the scrubbed code itself and an availability sentence.
Structure for a seven-minute visit
Reviewers give supplements minutes. Design entry points, not chapters:
supplement.zip
├── 00_INDEX.pdf # one page: what is where, mapped to paper claims
├── A_extended_results/ # grids continuing the body's table numbering
├── B_ablations/ # research question restated above each grid
├── C_failure_modes/ # taxonomy first, examples second
├── D_videos/
│ └── INDEX.txt # clip → claim → timestamp of the money moment
├── E_implementation/ # configs, preprocessing, hyperparameter tables
└── F_code/ # see iccv-artifact-evaluation for packaging bar
Two conventions repay their cost every review: continue the paper's numbering (Suppl. Table 9 follows body Table 8, so discussion-week comments are unambiguous), and put the single most persuasive qualitative result first in the video index — a reviewer who watches one clip should watch your best.
Failure gallery: the credibility purchase
An ICCV supplement with zero failure cases reads as curated, and vision reviewers are the world's most practiced detectors of cherry-picking. Budget a failure section with a two-level taxonomy (when it fails; why it fails there) and cross-reference it from the body's limitations paragraph. This is the cheapest credibility available at the venue: it costs one page of the archive and no page of the paper.
Recurring archive defects
- Dangling body pointers ("Suppl. G") after a late reorganization — grep the
final PDF for
Suppland resolve each against00_INDEX. - Config dumps contradicting the body's stated hyperparameters because one was updated post-freeze — regenerate both from one source of truth.
- Videos that only play in one player — test the archive on a machine that never touched your toolchain.
- A 300 MB gallery where 30 MB of curated depth would review better — size limits for 2025 were not verifiable (待核实), so design for a strict cap.
- Third-party assets carrying their authors' watermarks into your anonymous archive.
Reverify each cycle
- Whether the unified deadline persists in 2027 or the supplement regains its own date.
- Size caps and permitted formats (待核实 even for 2025).
- Any change to the do-not-cite-your-repo wording.
Output format
[Supplement verdict] ship / fix-first
[Parallel production] owner named · render slots reserved: yes/no
[Placement audit] decision-critical content in archive: none / <move list>
[Leak sweep] paths · video tags · configs · image metadata: clean?
[Navigation] index present · numbering continues body · best clip first
[Pointer resolution] body references resolving: n/m