混合餐盤:為什麼複合餐仍難倒多數 AI 追蹤器
我們對 1,200 份多成分餐盤做了基準測試。一旦菜中超過四種成分,領先者與落後者的差距迅速擴大。
上方標題與摘要已翻譯;正文目前為英文,我們正在進行在地化。
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