Concept: Calcifer – Visual Calorie Estimator

Use (deep) neural networks to directly model total calories in an image.

For a given food:

1. Take photos from various angles and in various conditions
2. Weigh just the food—no containers, labels, or packaging
3. Blend up the entire food until it is a consistent puree
4. Use a bomb calorimeter or similar to find the number of calories in a gram of the puree
5. Multiply by total number of grams—this becomes the target for a convnet regression problem

Repeat for a ton of foods. Also repeat for non-foods since food-ness is part of the prediction. (Not JUST calories.)

Possibly constrain the convnet into one sub-network per macro component (protein, fat, alcohol, carbohydrate)