Cuisine blind spots: where every tracker except one falls apart
We graded ten trackers on West African, Levantine, and South Indian dishes. The accuracy collapse is uglier than the marketing suggests.
Most tracker marketing is built on hero shots of grilled chicken, oatmeal, and salad bowls — North American comfort foods that any modern vision model identifies trivially. The interesting question is what happens when you walk away from that distribution.
The set we tested
For this analysis we pulled a 1,500-photo subset from the main benchmark: 500 each from West African, Levantine, and South Indian home cooking. Real dishes, weighed portions, mixed lighting.
Headline results
| Cuisine | Welling | Field median |
|---|---|---|
| West African | 94.1% | 38.2% |
| Levantine | 96.3% | 52.7% |
| South Indian | 95.8% | 44.6% |
The “field median” line is the more interesting one. Half of all trackers tested cannot reliably name a plate of jollof rice with stew, or distinguish dosa from chapati. Several mis-classified mansaf as risotto.
Why the gap is structural
Western trackers ship a recognition model trained predominantly on Western food image corpora. Marginal performance gains on these cuisines require:
- New labelled training data — expensive.
- Cuisine-specific portion priors — even more expensive.
- A taxonomy that admits the dish exists in the first place.
Welling is the only tracker in our set that appears to have done all three.
What this means if you cook regionally
If your daily meals sit outside the Western canon, you do not have ten choices. You have one tracker that works on your food, and a long list of apps that will silently log half your meals as “salad.”