技能 人工智能 WACV论文可复现性指南

WACV论文可复现性指南

v20260724
wacv-reproducibility
本指南旨在提高学术论文(尤其是涉及系统应用的论文)的科学可复现性。它要求作者建立详细的“配方账本”,记录所有报告指标的生成过程,包括数据分割、超参数、设备型号和环境约束,确保论文与实际代码/模型之间零漂移同步。
获取技能
343 次下载
概览

WACV Reproducibility

Use this to make a WACV result checkable — by a reviewer now and by you at the Round 2 resubmission. WACV's applications framing raises the bar in one direction (a systems claim must be reproducible as deployed), and the two-round model adds a second (paper and artifact must not drift between rounds). Facts are the WACV 2026/2027 cycles as read on 2026-07-09.

The recipe ledger

Keep one ledger that regenerates every reported number, so the body, the supplement, and the artifact cannot diverge:

Ledger entry Why WACV cares
Exact data splits and preprocessing Applications datasets are often custom; a hidden split invalidates a comparison
Seeds (and sessions/devices for field work) Reviewers distrust single hero runs
Hyperparameters per reported row Lets a reviewer see the comparison was matched
Device, power meter, and measurement method An applications latency/wattage claim is only reproducible if the rig is named
Baseline re-tuning under your constraint Proves the comparison was fair, not defaults-vs-yours
Script → figure/table mapping So a Round 2 reviewer confirms nothing changed silently

Constraint-aware reproducibility

An Applications-track claim ("2 W, sub-10-lux, on device D") is not reproducible from accuracy alone. Record how the constraint was measured — the meter, the device firmware, the ambient condition — so a reviewer or a future reader can reproduce the constraint, not just the metric. A number without its measurement rig is a claim, not evidence.

Repro smoke check before submission (and again before the R2 resubmission):
  1. Fresh checkout → run the pipeline for one reported row end to end.
  2. Confirm the produced number matches the paper within the stated spread.
  3. Diff the artifact's claims against the current paper's claims — zero drift allowed.
  4. Strip identity from the anonymous package (see wacv-artifact-evaluation).

Seed and session honesty

Report variance over seeds, and for deployed/field systems over repeated sessions or devices. Do not report the best of many runs as "the" result. If a gap sits within the spread, say so — an honest small margin survives review better than an inflated one that a reviewer's own reproduction contradicts.

Sync across the two rounds

The Revise-and-Resubmit lap is where reproducibility quietly breaks: authors change an experiment in the paper but not in the artifact, or vice versa. After every revision, re-run the smoke check and re-diff the artifact against the paper. A Round 2 reviewer re-reading a revised submission should find the package and the paper telling one story.

Reverify each cycle

  • Whether the current guidelines request a reproducibility statement or checklist.
  • Data-release and licensing rules for any dataset used as evidence.
  • Supplementary size/format caps that constrain what you can ship (待核实 for 2026).

Output format

[Recipe ledger] regenerates every reported number: yes/no
[Constraint rig] device/meter/condition recorded for systems claims: yes/no
[Seeds/sessions] variance reported honestly: yes/no
[Baselines] re-tuned under your constraint and logged: yes/no
[Round sync] artifact matches current paper (zero drift): yes/no
[Gap] <the number a reviewer could not currently reproduce>
信息
Category 人工智能
Name wacv-reproducibility
版本 v20260724
大小 3.67KB
更新时间 2026-07-29
语言