複合プレート:多成分の食事がいまだ AI トラッカーをつまずかせる理由
1,200 件の多成分プレートをベンチマーク。具材が 4 つを超えると、先頭と最下位の差が一気に広がる。
見出しと要約は翻訳済みです。本文は現在ローカライズ中のため、英語のまま掲載しています。
Single-item plates are easy. A chicken breast, a banana, a packaged protein bar, these are essentially classification problems with a tabular lookup at the end.
The hard cases are composite plates: a curry with rice and three sides, a Buddha bowl, a Sunday roast.
The 4-ingredient cliff
We sliced our 1,200 composite-plate sub-test by ingredient count. Up to four ingredients, the top five apps clustered within two composite points of each other. From five ingredients on, the field fans out.
By eight ingredients, Welling leads the next-best app by 14 composite points. Two of the bottom three apps simply refuse to log plates above six ingredients, they return a generic “mixed bowl” estimate that is almost always wrong.
What separates the leaders
Three behaviours show up in the top tier and nowhere else:
- Segmentation before identification. The model finds component boundaries first, then names each region. This avoids the failure mode of identifying a Buddha bowl as “salad” and calling it a day.
- Per-component portion estimation. Each segmented region gets its own gram estimate. Most apps still apply a global portion prior to the whole plate.
- Coverage of cuisine-specific composites. A katsu curry isn’t a curry plus a cutlet, the breading mass matters. Apps trained on Western cuisine alone get the macros wrong even if they identify all the components.
Take-home
If most of your meals are composite, bowls, plates, leftovers, pick from the top three. The field below that has not yet solved this problem.
Related: best tracker for vegan plates, best for restaurant dining, and our Welling vs MyFitnessPal comparison.
External references
- USDA FoodData Central, the underlying database most apps lean on.
- McCance & Widdowson’s composition of foods (UK)
- Japanese MEXT food composition tables