cvpr-topic-selection
GitHub辅助判断研究项目是否适合投递CVPR，评估视觉贡献核心性、各类型论文匹配度及拒稿风险。提供SOTA现实性检查清单，并指导将非核心视觉或理论工作路由至NeurIPS、ICCV等更合适 venue。
Trigger Scenarios
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill cvpr-topic-selection -g -y
SKILL.md
Frontmatter
{
"name": "cvpr-topic-selection",
"description": "Use when deciding whether a project belongs at CVPR or should route elsewhere, covering what counts as a vision contribution at the field's flagship, fit tests for methods, datasets, and application papers, realistic odds at 25% acceptance and 16k submissions, and routing to ICCV, ECCV, WACV, 3DV, NeurIPS, or a journal."
}
CVPR Topic Selection
CVPR is the largest venue in computer vision and one of the largest in all of science — 16,092 reviewed submissions and 4,090 acceptances in 2026. Size cuts both ways: almost any vision-adjacent topic has a reviewer pool there, and almost any weakness has a reviewer who has seen it a hundred times. This skill decides whether to feed the machine before other skills decide how.
The core question
Strip the engineering and ask: is the contribution a claim about visual data or visual computation? CVPR's 2026 program clustered exactly there — the largest areas were image/video synthesis and generation; vision+language and reasoning; multimodal learning; 3D from multi-view and sensors; and medical/biological vision (official program announcement). Contributions where vision is merely the demo domain — a generic optimizer tested on ImageNet, an ML theory result with a CIFAR table — historically route better to NeurIPS/ICML, where the reviewer pool evaluates the actual claim.
Fit tests by contribution type
| You have… | CVPR-shaped if… | Warning sign |
|---|---|---|
| A method/architecture | It solves a visual mechanism (geometry, temporal, pixels-to-structure), with benchmark wins + ablations | Gain vanishes under matched backbones |
| A dataset/benchmark | It unlocks a task the field cannot currently study, with baselines and analysis | "Bigger than the last one" is the whole pitch (and release is due at camera-ready) |
| A systems/efficiency result | Accuracy-per-FLOP frontier moves; CRF-style reporting is your friend | Speedup only on your hardware story |
| A vision-language model result | The visual grounding is the contribution | It's an LLM paper wearing an image encoder |
| An application (medical, agriculture, driving) | A general vision insight travels beyond the application | Domain novelty only → domain venue or WACV |
| Theory about vision | Predicts something checkable in experiments | Pure theory → NeurIPS/ICML/SIGGRAPH-adjacent |
The honesty checklist before committing a semester
- Leaderboard reality: are you within striking distance of the current SOTA on the benchmarks reviewers will demand, with the compute you actually have?
- Delta nameable: can you state, in one sentence, the mechanism that differs from
the three nearest papers? (If not yet, see
cvpr-related-workfirst.) - Ablatable: does the idea decompose into testable design decisions, or is it one entangled trick?
- Visual evidence exists: will qualitative results/figures show the improvement, or is it only a fourth-decimal metric story?
- Team can pay the process tax: November triple deadline, coauthor reviewer duties with desk-reject enforcement, a one-page January rebuttal — the process itself consumes a person-month.
Routing map
Contribution core → First-choice venue
──────────────────────────────────────────────────────────
Flagship vision method/benchmark → CVPR (Nov) — or ICCV/ECCV, same bar,
different months: pick by readiness date
Solid but not flagship-flashy; → WACV (applications-friendly CVF venue)
applications emphasis
3D/geometry-centric community → 3DV (also CVF-affiliated), or CVPR 3D areas
Learning theory / generic ML → NeurIPS / ICML / ICLR
Graphics-adjacent synthesis → SIGGRAPH (different review culture entirely)
Mature, extended, archival → TPAMI / IJCV (journal timelines, no rebuttal
sprint, room beyond 8 pages)
Early or niche idea → CVPR workshops (separate CFPs, lower stakes,
same audience walking past your poster)
CVPR vs. ICCV/ECCV is rarely a quality question — the bar is comparable and reviewer pools overlap — it is a calendar question: which deadline does your evidence mature for? Submitting a month early to the "bigger name" with a missing ablation is how teams donate a cycle.
Three worked verdicts (fictional projects)
- "We fine-tuned an open VLM on our agriculture dataset and accuracy rose 6 points." → Not CVPR-shaped yet. The contribution is domain data + recipe. Routes: WACV (applications) or a domain venue — unless analysis reveals a general insight about when VLM grounding fails, which could anchor a CVPR paper with broader experiments.
- "A test-time geometry constraint makes any monocular depth model temporally consistent, +X on three benchmarks, 2ms overhead." → CVPR-shaped. Visual mechanism, plug-in generality, ablatable, cheap to evaluate broadly; the risk to audit is baseline freshness.
- "A new loss improves classification on CIFAR/ImageNet, with a convergence theorem." → Split decision. As stated, it is an ML-methods paper (NeurIPS/ICML reviewers evaluate the theorem properly). It becomes CVPR-shaped only if the loss exploits something visual (spatial structure, augmentation geometry) and the evidence spans vision tasks beyond classification.
Scale realism
25.42% acceptance means the modal outcome for a competent paper is rejection, and tier outcomes concentrate attention further (in 2026, ~3–4% of the program presented orally). Choose CVPR when the upside justifies that variance: maximal audience (about 12,200 registrants in 2026), industrial visibility, and the strongest possible signal when a benchmark claim survives this particular gauntlet.
Main conference vs. CVPR workshops
The workshop program (separate CFPs, typically spring deadlines for a June conference) is a legitimate destination, not a consolation prize: new-task papers build their first community there, datasets get early adopters, and the audience walking past a workshop poster is the same 12,000-person crowd. Route to a workshop when the idea is promising but the main-conference evidence bar (leaderboard proximity, full ablations) is a cycle away — and note that workshop publication may interact with later dual-submission rules, so check both CFPs before using one as a stepping stone.
Reverify each cycle
- Current CFP topic list — areas are re-cut per edition (待核实 for 2027 until its CFP posts).
- Sibling-venue deadline calendar for the routing decision.
- Workshop CFPs, which appear months after the main-conference CFP.
- Acceptance-rate and program-shape statistics for the newest completed edition; the 16k/25% figures above are the 2026 snapshot, not a constant.
Output format
[Verdict] CVPR / sibling (which) / journal / workshop / not yet
[Core claim] <one sentence, visual-contribution phrasing>
[Fit evidence] leaderboard distance · nameable delta · ablatable · visual evidence
[Process tax] team can cover duties + rebuttal week: yes/no
[Route if not CVPR] <venue + verified deadline>
[Ripeness gap] <what must exist before committing>
Version History
- 9f86f09 Current 2026-07-19 14:42
