This repository contains the code, the results and documents used for our master thesis project in Engineering. The end result is a furniture setup guide for iOS that utilizes Augmented Reality, Object Detection and Object Recognition. For further in depth explanation about the project and its goal, read the report in Documents/Report.
To train and generate the CoreML model for the network, using our model built from scratch, enter the Trainer directory. First you need to install the following dependencies:
- Python 2.7
- Pillow
- tensorflow 1.5.0
- keras version 2.1.3
- numpy
- matplotlib
- scikit-learn
- tfcoreml
These can all be installed by going to the Training folder and running
make installTo then scale the images and also generate artificiall data, run:
make generate-dataAfter generating the data, you can train a keras model with the following command:
make trainOne then has to convert the keras .h5 model to a .mlmodel file, which is then useable in the Xcode project. This is done by running:
make convertTo run the three aforementioned commands in one go, just write:
make setupTo try out the transfer learning method, enter the FeatureTraining. First you need to run the python script resize.py which scales the original images to correct image size. Then you can train the net by running transferLearning.py, which will generate a .h5 file. Currently no direct way to convert this to a .mlmodel file. At the moment, you have to make use of the convertKerasToCoreML.py script in Trainer/Model folder.
To be able to compile the project xcode, you have to have a .mlmodel file named FurnitureNet.mlmodel in the folder Trainer/Model.
A UML Class diagram showing the flow of the project can be seen bellow.
Further interesting links, the report and other goodie documents can be fold in the Documents folder.
Created by Jan Svensson(iSadist) and Jonatan Atles(Oden-Allfader)