技能 人工智能 机器学习研究成果物评估

机器学习研究成果物评估

v20260724
iccv-artifact-evaluation
本指南详细介绍了顶级机器学习会议的科研成果物(代码、模型、数据)的最佳发布和维护实践。内容覆盖了匿名评审时代码的打包规范和公开发布的周期管理,旨在确保研究的高度可复现性、可追溯性及许可的清晰性。
获取技能
279 次下载
概览

ICCV Artifact Evaluation

ICCV grants no artifact badges; the judgment is social and it is severe, because the venue's own history sets the bar. ICCV papers like Mask R-CNN (2017), Swin Transformer (2021), ControlNet (2023), and Segment Anything (2023) are cited as much for their released artifacts as for their PDFs — at this venue, the release is part of the contribution's afterlife. Package for two moments: the anonymous review in spring, and the public release whose natural deadline is the October conference.

Review-time: the sealed archive

The 2025 rules shape review packaging in two specific ways. First, do not cite or link your public codebase — anonymity guidance says to write that code "will be made publicly available" instead. Second, paper and supplement land on the same day, so the code package cannot be an afterthought of a post-deadline week (iccv-supplementary). Inside the archive:

  • Ship source, configs, and exact environment pins; leave the weights out unless small — describe which checkpoints the release will include.
  • Scrub identity the way archives actually leak: git metadata, usernames in paths, experiment-tracker entity names, cluster hostnames in launcher scripts.
  • Include five-sample inference with expected outputs, so a reviewer can smoke- test without your datasets.
# Anonymized review package from a live repo, reproducibly
git archive HEAD -o /tmp/pkg.tar && mkdir -p /tmp/pkg && tar -xf /tmp/pkg.tar -C /tmp/pkg
cd /tmp/pkg
rm -rf .github wandb outputs
grep -rliE "$(git config user.name | awk '{print $1}')|$(hostname -s)" . | head   # identity residue
python -m pip freeze > requirements.lock                                          # pin exactly what ran
printf 'table→command map:\n  Tab.1: bash run/eval_main.sh\n' > REPRODUCE.md
zip -qr ../iccv_code.zip .

REPRODUCE.md — one command per paper table/figure — is the highest-value file per byte in the archive; it converts "trust the tables" into "verify one row."

The four-month release runway

ICCV's calendar gives artifact work a gift no November-deadline venue offers: decisions in late June, conference in mid-October. Plan the runway backward from conference week, when attention peaks:

Weeks after decision Release milestone
0–2 License chosen; hosting decided; repo made public with paper, citation, and "weights coming "
2–6 Weights uploaded with checksums; inference path verified from a clean machine outside your network
6–10 Training/eval reproduction path documented; issues from early adopters triaged
by conference Everything in the paper's availability statement is live; project page links CVF open access

A repo that goes live during conference week arrives exactly when thousands of people are reading the paper; a repo that goes live three months later arrives after the field has moved to reproducing someone else's implementation of your method — which then becomes the cited one.

Decisions to make on purpose

  • License, split by asset: code (permissive vs copyleft), weights (research- only vs unrestricted — decide who may fine-tune and redistribute), data (can you even relicense what you scraped?). "No license" reads as "do not use" to every industrial lab.
  • Checkpoint provenance: name checkpoints for the table rows they reproduce and record the code commit that evaluates each to the printed number.
  • Data documentation: sources, filtering, consent posture for human imagery. Vision datasets get audited years later; write the provenance section while the facts are fresh.
  • Generative-model posture: for synthesis work, decide what ships (full weights, adapters, inference-only) and write the misuse considerations yourself before a journalist does.
  • Benchmark freeze: if you release evaluation code others will report numbers from, version it and changelog every metric fix — silent corrections fork leaderboards.

The infrastructure test

Before calling the release done, run the test the field will run:

  1. Fresh cloud machine, no lab credentials: does git clone → env setup → five-image inference work inside an hour?
  2. Do the printed numbers reproduce? State the tolerance you observed (iccv-reproducibility); a ±0.2 note preempts a hundred issues.
  3. Is there exactly one entry point per use case (inference / train / eval), each documented in under a screen of README?
  4. Would a person with only the checkpoint and the eval script get your Table 1 row? If not, what unstated asset is missing — and can it be shipped?

Maintenance as reputation

The two-year gap to the next ICCV means your artifact is the state of the record for a long time. Budget light maintenance: pin dependencies against rot, tag the camera-ready state (v1.0-iccv<year>), answer the first month's reproduction issues (usually environment drift), and keep a short log of community-confirmed reproductions. The venue's landmark releases earned their citation counts in this phase, not at upload.

Reverify each cycle

  • The availability-statement and repo-citation wording in the current Author Guidelines.
  • Supplement size limits that constrain review-time code (待核实 for 2025).
  • Any artifact checklist or badge program the 2027 chairs may introduce.

Output format

[Phase] review-archive / release-runway / maintenance
[Archive] identity-clean · pinned · smoke-testable · REPRODUCE.md maps n/m tables
[Runway] weeks to conference: <n>; milestones on track: yes/no
[Licenses] code/weights/data each decided: yes/no
[Infrastructure test] clean-machine hour test: pass/fail at <step>
[Debt] <unshipped promises from the availability statement>
信息
Category 人工智能
Name iccv-artifact-evaluation
版本 v20260724
大小 6.02KB
更新时间 2026-07-28
语言