mikeczech / seefood Goto Github PK

As an end user I want to easily take pictures of food on iOS so that I can track my eating habits.

Given the user has pressed the button for entering the camera mode
When the button is pressed again
Then a picture is taken and displayed to the user

Subtasks

• create camera view that shows a preview to the user
• create button event for taking a picture
• display picture in another view

As an end user I want to track my eating habits so that I can see if I live up to my healthy nutrition plan.

Given a picture was taken using the app
When the save-button was pressed
Then the app saves the picture, the label, and the amount of calories AND displays the app's main view

Given multiple pictures of food were taken using the app
When the app's main view is opened
Then the app shows a list of all consumed food items including a label and the amount of calories for each item

Note that this issue also demands local persistence of the tracked food items.

This issue depends on #1, #2, and #3

As an end user I want to quickly estimate calories from images of food so that I can determine it it fits into my healthy eating plan.

Given the camera displays a single item of food
When the button for taking a picture is pressed
Then an estimation for the amount of calories of that item appears

Note that the estimation of calories can be arbitrary (e.g. random).

This issue depends on #2

As an end-user I want to directly use the model for label prediction from the app so that I can precisely track my meals.

Given the app is opened
When pictures of food are taken using the app
Then all displayed food labels come from a model that is locally deployed on the iOS device

In contast to using a web service for model serving, a local deployment ensures functionality without a stable internet connection and simplifies the overall architecture in this early stage of development.

This story depends on #3

As an end-user I want to directly use the model for calorie estimations from the app so that I can precisely track my calorie intake.

Given the app is opened
When pictures of food are taken using the app
Then all displayed calorie estimations come from a model

This story depends on #1

As a data scientist I want to evaluate the usage of simple NNs for label prediction so that I gain further insights for getting an easily deployable, but also production-ready model.

Given Sparkrecipes dataset, SqueezeNet Image embeddings, and a set of food labels
When training a NN to predict food labels
Hypothesis the NN predicts the labels reliably ( ~95% accuracy)

Note that this issue also includes the definition of a fixed set of labels. Those labels can be very abstract in order to simplify the task (e.g. "Soup" instead of "Tom Yum Soup")

As a data scientist I want to operate on clean data so that I can be sure that my results are correct.

Subtasks

• remove images

As an end user I want to have a label for each picture I have taken so that I can track my eating habits.

Given the camera displays a single item of food
When the button for taking a picture is pressed
Then a label for that item appears

Note that the label can be arbitrary (e.g. random) and the specific set of possible labels will be defined in a different issue.

This issue depends on #2

As a data scientist I want to evaluate the usage of simple NNs for calorie estimations against the baseline and other models so that I gain further insights for getting an easily deployable, but also production-ready model.

Given Sparkrecipes dataset and SqueezeNet Image embeddings
When Training a NN to predict calories
Hypothesis predictions are superior to the baseline and almost identical to other models like XGBoost / linear regression

As a data scientist I want to make sure that my Test-Train split does not contain a bias so that I make sure that my model is correct.