My own take on estimating calories based on an object detection app based on YOLOv8. Some software used include Android Studio (with Kotlin), Google Colab, and Python. Dataset obtained from Roboflow. More info on the bottom for helpful repositories.
The project right now only detects the food. The calorie predictions will be done by cross-referencing the predicted file using a dataset and displaying it through the app's GUI. You may even fork this and extend that functionality if you want.
The model used was trained with 100 epochs (I later adjusted the epochs to 50 for easier use on low-end devices, but this was not used for in the final model). There are a total of 22 classes based on the dataset used. These classes can be found on the data.yaml file. Below are some of the predictions of the TFlite model.
You may download the file here and try it out for yourself.
- Expand the dataset (at least 500 images per category).
- Use a faster device to model (to train 100 epochs took my computer 20 hours).
- Ultralytics
- Other Python modules