ACM MM Artifact Evaluation
Use this to turn an ACM Multimedia project's code, models, media, and data into the right
artifact for the right track. ACM MM has a track economy around artifacts, and the choice
determines blinding, format, and what reviewers judge.
Which track is the artifact?
| Artifact is primarily... |
Route to |
Blinding |
Judged on |
| A reusable software system/framework |
Open Source Software Competition |
Single-blind |
Adoption, quality, license, docs |
| A new dataset/benchmark |
Dataset track |
Single-blind |
Scale, quality, ethics, usefulness |
| A reproduction of published results |
Reproducibility track |
Single-blind |
Whether results rebuild; ACM badges |
| Supporting evidence for a method paper |
Main-track supplement |
Double-blind |
Whether it backs the paper's claims |
The named single-blind tracks exist because the artifact's identity cannot be hidden; a
main-track method paper's artifact, by contrast, must be anonymous through review.
Two artifacts, two audiences
Plan both from the start:
- Anonymous review artifact — what reviewers see during double-blind review: an
anonymized repository, an anonymous data mirror, stripped media metadata, and a README that
reveals no author identity.
- Public release artifact — what ships at/after camera-ready: the de-anonymized
repository, a permanent archive (DOI), the license, and the final dataset/model.
review/ -> anonymous repo, anon data mirror, no names in code/media, run instructions
release/ -> public repo + DOI, LICENSE, model weights, dataset card, citation
Open Source Software Competition
- The bar is a system others will use: clear install, documentation, examples, an
OSI-approved license, and evidence of quality or adoption.
- Reference models and reproducible examples matter more than a single benchmark number —
this is the lane exemplified by community frameworks and portable libraries.
Dataset track
- Ship a dataset card: collection method, size, splits, license, consent, and known
biases or limitations.
- Address ethics and rights explicitly, especially for user-generated or scraped media; a
dataset a reviewer cannot legally use is not a contribution.
Licensing and rights decisions
- Choose a code license (permissive vs. copyleft) and a data license separately; they are
not the same choice.
- For media, confirm you have the right to redistribute; where you cannot, provide a
retrieval script or agreement path instead of the raw files.
- Record third-party asset licenses so the release is clean.
Ethics and consent for media artifacts
Multimedia artifacts carry people's faces, voices, and content, so the ethics review is not a
formality:
- Consent and rights — confirm you may redistribute the media; user-generated content often
cannot be re-hosted, so ship a retrieval script or agreement path instead.
- Privacy — remove or justify identifiable individuals who did not consent; a dataset of
scraped faces is a rejection risk regardless of its scale.
- Documentation — a dataset card that states collection method, consent, license, and known
biases is part of the contribution, not paperwork.
Timeline: review artifact, then release
before paper deadline: anonymous review artifact ready (repo + data mirror, no identity)
during review: reviewers/AC access the anonymous artifact
on acceptance: build the public release (de-anonymized repo + DOI + license)
by camera-ready: release replaces the anonymous mirror; dataset/model final
Plan the public release early even though it ships late — a scramble at camera-ready is how
projects end up with a broken anonymous link and no working public archive.
Output format
[Track] Open Source / Dataset / Reproducibility / main-track supplement
[Blinding] correct for track / mismatch
[Review artifact] anonymous + runnable / gaps: <list>
[Release artifact] archived + licensed / gaps: <list>
[Rights] code+data+media licenses set / open questions: <list>
[Top fixes] <ordered>
1---2name: acmmm-artifact-evaluation3description: Use when packaging code, models, datasets, or media as ACM MM (ACM Multimedia) artifacts — building the anonymous review package versus the public release, and choosing between the Open Source Software Competition, the Dataset track, the Reproducibility track, and main-track supplementary evidence, each with its own blinding and expectations.4---56# ACM MM Artifact Evaluation78Use this to turn an ACM Multimedia project's code, models, media, and data into the *right*9artifact for the *right* track. ACM MM has a track economy around artifacts, and the choice10determines blinding, format, and what reviewers judge.1112## Which track is the artifact?1314| Artifact is primarily... | Route to | Blinding | Judged on |15|---|---|---|---|16| A reusable software system/framework | Open Source Software Competition | Single-blind | Adoption, quality, license, docs |17| A new dataset/benchmark | Dataset track | Single-blind | Scale, quality, ethics, usefulness |18| A reproduction of published results | Reproducibility track | Single-blind | Whether results rebuild; ACM badges |19| Supporting evidence for a method paper | Main-track supplement | Double-blind | Whether it backs the paper's claims |2021The named single-blind tracks exist *because* the artifact's identity cannot be hidden; a22main-track method paper's artifact, by contrast, must be **anonymous** through review.2324## Two artifacts, two audiences2526Plan both from the start:2728- **Anonymous review artifact** — what reviewers see during double-blind review: an29 anonymized repository, an anonymous data mirror, stripped media metadata, and a README that30 reveals no author identity.31- **Public release artifact** — what ships at/after camera-ready: the de-anonymized32 repository, a permanent archive (DOI), the license, and the final dataset/model.3334```text35review/ -> anonymous repo, anon data mirror, no names in code/media, run instructions36release/ -> public repo + DOI, LICENSE, model weights, dataset card, citation37```3839## Open Source Software Competition4041- The bar is a system others will *use*: clear install, documentation, examples, an42 OSI-approved license, and evidence of quality or adoption.43- Reference models and reproducible examples matter more than a single benchmark number —44 this is the lane exemplified by community frameworks and portable libraries.4546## Dataset track4748- Ship a **dataset card**: collection method, size, splits, license, consent, and known49 biases or limitations.50- Address ethics and rights explicitly, especially for user-generated or scraped media; a51 dataset a reviewer cannot legally use is not a contribution.5253## Licensing and rights decisions5455- Choose a code license (permissive vs. copyleft) and a **data** license separately; they are56 not the same choice.57- For media, confirm you have the right to redistribute; where you cannot, provide a58 retrieval script or agreement path instead of the raw files.59- Record third-party asset licenses so the release is clean.6061## Ethics and consent for media artifacts6263Multimedia artifacts carry people's faces, voices, and content, so the ethics review is not a64formality:6566- **Consent and rights** — confirm you may redistribute the media; user-generated content often67 cannot be re-hosted, so ship a retrieval script or agreement path instead.68- **Privacy** — remove or justify identifiable individuals who did not consent; a dataset of69 scraped faces is a rejection risk regardless of its scale.70- **Documentation** — a dataset card that states collection method, consent, license, and known71 biases is part of the contribution, not paperwork.7273## Timeline: review artifact, then release7475```text76before paper deadline: anonymous review artifact ready (repo + data mirror, no identity)77during review: reviewers/AC access the anonymous artifact78on acceptance: build the public release (de-anonymized repo + DOI + license)79by camera-ready: release replaces the anonymous mirror; dataset/model final80```8182Plan the public release early even though it ships late — a scramble at camera-ready is how83projects end up with a broken anonymous link and no working public archive.8485## Output format8687```text88[Track] Open Source / Dataset / Reproducibility / main-track supplement89[Blinding] correct for track / mismatch90[Review artifact] anonymous + runnable / gaps: <list>91[Release artifact] archived + licensed / gaps: <list>92[Rights] code+data+media licenses set / open questions: <list>93[Top fixes] <ordered>94```
brycewang-stanford