[Week 5 — SeeFood]

Theme: Food Calorie Estimation

Team Members: Okan ALAN Gökberk Şahin Emre Yazıcı

Perfectly balanced, as all things should be.

This week we tried to train an object detection model on our dataset using Tensorflow Object Detection API. We used the Faster R-CNN Inception model from the Tensorflow models GitHub repository. Model achieved the loss of 1.3 at 250th step so we stopped the training for testing. Here are results of the test images:

Detects the coin but not the banana

It detects the coin quite accurately but somehow it won’t detect the banana. We double checked our work but couldn’t find a point that we screw up. I think our model is so stingy that it only recognizes coins. Here are the results for Turkish coins:

Next week we’ll try to solve this problem and we’ll move on to segment this images using the GrabCut algorithm. When we’re done with these whole detection things we can actually do the cool stuff that we look forward to. We have some ideas about how we’re gonna estimate the calories besides using the boring formulas. We think that if we feed forward the bounding boxes of the foods and coins, number of pixels after the GrabCut and the real volume as the label we can build a Multi-Layer Perceptron that will learn the mapping between these features and the label.

Thanks for reading and see you next year!