brycewang-stanford ACM MM Related Work
Use this to position an ACM Multimedia paper against a literature that is unusually spread out: a cross-modal contribution touches several single-modality communities plus the multimedia venues that combine them.
The five multimedia shelves
A strong ACM MM related-work section shows command of all shelves the contribution touches, not just the author's home community:
- Vision — the visual side (CVPR/ICCV/ECCV) your method builds on or competes with.
- Audio/speech and music — the acoustic side (ICASSP, INTERSPEECH, ISMIR) if a stream is audio.
- Language — the text side (ACL/EMNLP) if captions, transcripts, or descriptions matter.
- HCI / QoE / human-centric — perception, engagement, and interaction (CHI, QoMEX) when the claim is subjective.
- Multimedia proper — ACM MM, ICMR, MMSys, and TOMM, where these threads are combined.
Coverage vs. venue-hygiene table
| Task | What to do | Failure it prevents |
|---|---|---|
| Cover each modality you use | Cite the current best single-modality work per stream | "You ignored the vision literature" |
| Cite the fusion lineage | Trace the cross-modal line your method extends | "No delta over existing fusion" |
| Verify the venue string | Confirm each "ACM MM" cite on dblp conf/mm |
Misattributing an ICMR/CVPR paper to ACM MM |
| Handle concurrency | Note contemporaneous arXiv work honestly | "You missed / overclaimed novelty vs. X" |
| Keep it blind | Cite your own prior work in third person | Double-blind violation |
Positioning, not listing
Do not enumerate. For each closest neighbor, state in one clause what it did and in one clause what your paper adds — and make sure the delta is cross-modal, since "we swap a better encoder" is a single-modality delta a reviewer will discount.
[Neighbor] Late-fusion audio-visual highlight scoring (VenueYear).
[Their move] Average per-modality scores.
[Our delta] Score the timing DISAGREEMENT between streams — a signal averaging cannot represent.
Concurrency under arXiv speed
Multimedia sub-areas move fast and preprint heavily. Acknowledge genuinely concurrent work (roughly same-window preprints) as concurrent rather than prior, do not claim to beat a method you did not run, and do not silently drop a close preprint — a reviewer who knows it reads the omission as evasion.
Double-blind citation hygiene
- Cite your own earlier papers in the third person ("Prior work [X] showed..."), never "our previous paper."
- Avoid anonymity leaks through a dataset, system name, or repository that only your group uses.
- The single-blind tracks (Reproducibility, Open Source Software, Dataset) relax this — but the main track and Brave New Ideas do not.
Building the section in passes
Pass 1: list the modalities and sub-fields your contribution touches (the shelves).
Pass 2: for each shelf, cite the current strongest 2-3 works you build on or beat.
Pass 3: trace the FUSION lineage — the cross-modal line your method extends — as its own thread.
Pass 4: add same-window arXiv work as concurrent, not prior.
Pass 5: verify every "ACM MM" cite on dblp; convert your self-cites to third person.
A reader should finish the section able to name which prior fusion approach you improve on and why the single-modality shelves are covered but not the whole story.
The multimedia-specific omission risk
Because a cross-modal paper sits between communities, the dangerous omission is usually the other community's closest work — a vision-trained author who misses the audio or IR paper that already did half the job. Before submitting, ask a question from each shelf's perspective: "what would an audio reviewer, an IR reviewer, and a systems reviewer each say I missed?" That triage catches the omissions that sink cross-area papers.
Venue-verification pass
Before submission, spot-check every citation that claims an ACM MM placement against dblp's
conf/mm edition record. The common traps: ICMR papers cited as ACM MM, TOMM journal
articles cited as the conference, and CVPR/ICCV vision papers cited as multimedia. Fix the
venue string or the claim that rests on it.
Output format
[Shelf coverage] vision/audio/language/HCI-QoE/multimedia — <covered / gaps>
[Fusion lineage] traced / missing
[Deltas] cross-modal and specific / single-modality or vague: <list>
[Concurrency] handled / risky omissions: <list>
[Blindness] clean / leaks: <list>
[Venue hygiene] verified / suspect cites: <list>