Cvpr workflow

CVPR Workflow

CVPR is an annual machine with one intake: in the 2026 cycle, abstracts November 6, papers November 13, supplements November 20, rebuttal late January, decisions February 20, conference June 3–7 in Denver. This skill turns that spine into a project plan with owners, because CVPR deadline weeks are lost to coordination failures more often than to science failures.

Campaign calendar (2026-cycle anchors)

When	Milestone	Owner discipline
T–10 weeks (early Sep)	Freeze scope: claims, benchmarks, baselines	PI + first author sign one paragraph
T–8 weeks	Experiment matrix running; recipe ledger live	Every table row has a config
T–6 weeks	First full draft; flip-test the figures	Co-authors annotate, don't rewrite
T–4 weeks (early Oct)	OpenReview profiles audited for every author	One named owner; new coauthors close the list
T–1 week (Nov 6)	Abstract registered; profiles valid	Desk-reject exposure if missed
Nov 13	Paper + CRF submitted	Freeze 24h early; upload before the panic hour
Nov 20	Supplement uploaded	Videos/code packaged per cvpr-supplementary
Dec – Jan 12	Coauthors deliver their assigned reviews	Late reviews endanger your paper
Jan 22–29	Reviews land; one-page rebuttal	Pre-book this week; nobody travels
Jan 30 – Feb 5	Discussion (no author access)	Resist mailing the AC
Feb 20	Decisions	Both branches pre-planned
Spring (待核实)	Camera-ready + dataset release due	See cvpr-camera-ready
Jun 3–7	Conference (workshops 3–4, main 5–7 in 2026)	Poster/talk/travel logistics

The two windows people fail to defend

Deadline fortnight (early November). The supplement trailing the paper by a week tempts teams to leave packaging "for week two" — then week two evaporates into video rendering and code scrubbing. Rule: the supplement plan (what goes in, who builds it) is written before the paper deadline; week two only executes it.

Rebuttal week (late January). Seven days, one page, all reviewers at once. Assign roles when reviews drop, not after a day of group grief:

# rebuttal-roles.yaml — filled within 12 hours of review release
reader:        # summarizes all reviews neutrally, one page, no rebuttals yet
runner:        # owns the 1-2 small requested experiments; hard stop at day 5
writer:        # owns the single page; everyone else feeds bullets
adjudicator:   # PI decides which objections get space, which get one line
deadline:      # internal freeze 24h before OpenReview close

Reviewer duty as a planned workload

Submitting enrolls qualifying authors into a 25k-reviewer pool, and undelivered or irresponsible reviews can desk-reject every paper the offender authored (verified 2026 enforcement). Treat December–January reviewing as a scheduled deliverable: at submission time, list which coauthors are likely reviewers, block their calendar for review weeks, and have the group's senior authors sanity-check first-time reviewers' drafts against the official reviewer guidelines.

Multi-paper coordination

Most labs submit several papers into the same November wall, which multiplies the process risks nonlinearly:

• Shared coauthors, shared exposure: one person's undelivered January review endangers every paper they touch; the duty ledger must be lab-level, not paper-level.
• Conflict hygiene: OpenReview conflicts must be complete across all submissions; a stale institution on one profile pollutes assignments for the whole batch.
• Compute contention: rebuttal-week reserve GPU allocations collide when three papers each promised "small requested experiments"; pre-negotiate priority order.
• Internal review: schedule lab-mates' cross-reads at T–10, not T–2 — at T–2 everyone is defending their own draft.

The resubmission ladder

At a 25.42% acceptance rate, the rational plan covers rejection before it happens. The CV calendar offers a clean ladder — CVPR (Nov) → ICCV (spring deadline, odd years) / ECCV (spring deadline, even years) → WACV/3DV or back to CVPR. Rules for climbing:

• Bank the reviews: fix every factual objection before the next deadline; reviewer pools overlap across CVF venues and repeat objections read badly (triage method in cvpr-review-process).
• Mind the dual-submission windows: the CVPR attestation covered Nov 13 – Feb 20 in 2026; the next venue's window must not overlap a paper still under CVPR review.
• Re-check freshness: three months of arXiv can obsolete a comparison table; budget a baseline-refresh pass, not just an edit pass.

Deadline fortnight, day by day

The last two weeks compress badly when run ad hoc. A schedule that has survived contact with real teams:

• T–14 to T–10: results freeze for the main table. New numbers after this go to the supplement or the rebuttal, not the body.
• T–10 to T–7: full-draft pass; run the figure flip test with an outsider; final baseline-freshness check against the leaderboard.
• T–7 (abstract day): register the abstract, verify every profile renders as complete when logged out of the authors' accounts, lock the author list.
• T–6 to T–3: claim-calibration pass (cvpr-writing-style), bibliography venue verification (cvpr-related-work), CRF mandatory sections drafted.
• T–2: full mechanical sweep from cvpr-submission; internal freeze.
• T–1: upload; re-download; cold read. Nobody edits after upload day.
• T to T+7 (supplement week): execute — do not design — the supplement plan; identity-scan code and videos; upload at T+5, not T+7, to survive a failed render.

Cycle-status note

As of 2026-07-08, CVPR 2026 has concluded and the CVPR 2027 cycle (Seattle, reported around June 20–24, 2027) had published no CFP or deadlines. Every date above is a 2026-cycle anchor for planning shape only — reopen cvpr.thecvf.com the moment your team commits, because the deadline chain (and things like the CRF) can change between editions.

Output format

[Campaign stage] scoping / building / deadline-fortnight / in-review / rebuttal / decided
[Calendar] next 3 milestones with dates and owners
[Duty ledger] reviewing coauthors + calendar blocked: yes/no
[Risk register] top 3 (missed-profile, supplement drift, rebuttal roles, …)
[Plan B] next venue + its verified deadline
[Cycle check] current-year pages reopened on: <date>