acmmm-experiments

ACM MM Experiments

Use this to make an ACM Multimedia paper's evidence match its claim. The reviewer's implicit questions are: does it work, does the cross-modal part cause the gain, when does it fail, and — if the target is perceptual — do people actually prefer it.

The four questions and how to answer them

The second row is what separates an ACM MM experiment section from a single-modality one: if removing a modality does not move the result, the paper is not really cross-modal.

Matched baselines

• Compare against the strongest existing method, re-run under your data and preprocessing where feasible, not a weakened reimplementation.
• Include a late-fusion / naive-concatenation baseline so the reader sees what the fancy fusion buys over the obvious one.
• Hold everything but the mechanism fixed: same backbone, same features, same training budget, so the delta is attributable.

Ablations that isolate the cross-modal claim

Full model .................... reference
- audio stream ................ tests whether audio carries signal
- text/caption stream ......... tests whether language carries signal
- alignment / fusion module ... replaced by concatenation: tests the MECHANISM
- synchronization assumption .. shuffled timing: tests whether cross-modal timing matters

Report each ablation with the same metric and variance as the headline, and state which term carries most of the gain — reviewers reward a paper that can point to why it works.