This demo app allows users to quickly obtain calorie information using their camera. The object detection model utilizes a MobileNetV2 SSD architecture that was trained using transfer learning on 15 unique food classes from the Open Images v4 dataset.

Note: this app was built and modified from the TensorFlow Lite Object Detection Android Demo.

To build the pretrained demo in Android Studio, select "Open an existing Android Studio project" and navigate to the Food.AI/Food.AI directory. Then, connect a device and press 'run'.

Tip: to view additional details (e.g. detection confidence, inference time, etc.) when detecting foods, press a volume key.

• Create a directory in Google Drive called food_detection.

• Add the training dataset and label_map.pbtxt to food_detection.

• Open FoodAI_train.ipynb and follow the notebook instructions.

• To use the newly trained model, download food_detect.tflite from model_checkpoints/tflite_model/ and move it to the assets folder in Android Studio. It should replace the existing pretrained model.

• Create a directory in Google Drive called food_detection.

• Use OIDv4 ToolKit to download images and bounding box annotations for the desired classes.

• Change the classes in OIDv4_ToolKit/classes.txt accordingly. Then, zip the OIDv4_ToolKit folder and upload it to food_detection.

• Modify label_map.pbtxt to match the custom classes and upload it tofood_detection.

• When generating the TFRecords, set the flags to point to the location of the dataset.

• Edit the number of classes in the model configuration file (s = re.sub('90', 'NUM_CUSTOM_CLASSES', s)).

• Download food_detect.tflite from model_checkpoints/tflite_model/ and move it to the assets folder in Android Studio. It should replace the existing pretrained model.

• Modify food_labelmap.txt accordingly. Make sure to keep ??? as the first line.

• Modify calorie_info.txt to reflect the custom classes