Replicate this work with any data to create an art-recommendation HoloLens application.
This project shows an application for augmented reality in museums. We provide all the steps necessary to write an art-recommendation application for the Microsoft HoloLens. This project makes use of data from The Metropolitan Museum of Art (The MET) Collection API, but the code is written in such a way that it can be adapted to other data sources. The project consists of two major components, namely the backend (for data processing, image similarity search, and web app endpoint deployment) and the frontend (Unity application and visualization for the Hololens), in which we build off of this tutorial. We step through the components and explain how to replicate the work in this document.
Key Topics: Microsoft HoloLens, Unity, Computer Vision, Docker, Python, C#
These are the type of results you will have by the end of the guide. The first .gif demonstrates the current UI. The other videos illustrate the core components but at earlier stages of development.
Full functionality with current UI. See the video on YouTube here.
Perform an Air tap with the HoloLens to draw the information panels.
Information panels are spawned for each Air tap based on what you are looking at. Here you see the panels overlayed in the real-world. Note that the panels have a new UI, which is depicted in the GIF at the top.
Here is a closer look at an information panel. Sorted left to right, the small images represent the closest-looking images in your dataset (in our case, a portion of The MET data). Clicking on the small images brings up the information for that particular art piece. These 5 similar images came from searching the dataset with an image of myself, which is why the recommended art pieces are human depictions. (Ignore the white text behind the title; this was for testing purposes.)
Set up a virtual environment for local development before deploying to an Azure VM. These commands work on a Linux/Unix machine, but similar commands will work for other devices. If new to using virtualenv, check out their documentation here.
# first clone the repository
git clone <repository_url>
# navigate the the repository
cd path/to/the/repository
# create the virtual environment in a folder named venv
python3 -m venv venv
# start using the environment
source venv/bin/activate
# updgrade pip and install dependencies
pip install --upgrade pip
pip install -r main/requirements.txt
# install kernel for jupyter notebook
# this allows you to use this virtualenv in the notebooks
ipython kernel install --user --name=arart
# to deactivate the environment after use
deactivate
We use a virtualenv like this for local testing, but we use Docker containers for deployment (which is described later). virtualenv is simply faster when working on your local machine, but Docker makes it easy to deploy on an Azure VM. Because all dependencies are listed in the requirements.txt file, it's easy to go back and forth between virtualenv and Docker.
If using data from a source other than The MET, this section will require the most code on your end. We do our best to outline the necessary steps in the iPython notebooks. The final output will be a folder of images with a .csv file with fields of textual informtion.
Make sure you have Jupyter Notebook installed and set up before proceeding. Then run the following commands to start the Jupyter notebook:
# start the notebook in the terminal
jupyter notebook
# navigate to the specified address in a browser
# in the web GUI
click Kernel -> Change kernel -> arart
Note if doing this on an Ubuntu machine, you will need to install some dependencies for
import cv2to work.
# these commands should be run in the terminal before running the notebooks
sudo apt update
sudo apt install -y libsm6 libxext6
sudo apt-get install libsm6 libxrender1 libfontconfig1
-
0_CreateDatasetFromMetAPI.ipynb
- This notebook queries the MET API for a specified number of times to create a .csv file. Each row in the .csv file if for a different object in the museum. The columns are for each field in the dataset. The following image shows what the .csv should look like. Not all the entries will be complete, such as the image URL for objectID 36. We simply ignore these fields for this product.
-
- This notebook will use the .csv file created from CreateDatasetFromMetAPI.ipynb to download the images to an
imagesfolder. However, this may not be needed depending on the dataset used. If you have access to the images, you will just have to format aimagesfolder to have entries with<objectid>.jpg. We ignore the objects in the dataset that don't have an associated image.
- This notebook will use the .csv file created from CreateDatasetFromMetAPI.ipynb to download the images to an
-
2_CreateFeaturesDictionary.ipynb
- We use a pre-trained convolutional network, ResNet 18, trained on ImageNet for image similarity. ImageNet has 1000 categories, so this pre-trained network must learn relevant features to distinguish items from each other. From this network, we can extract the last embedding layer before the final softmax prediction. This leaves us with a 1000 dimensional vector for a given input image. These vectors can then be compared with L2 distance to get a metric for image similarity. Learn more about image search here. A dictionay keyed by objectID and with values of the vectors is serailized and stored to the pickle file,
features_dict.pickle. The code for the network that we use can be found infeature_extractor.py, that makes use of PyTorch. If wishing to speed up the process with a GPU, you can modify the code in a similar way to explained here.
- We use a pre-trained convolutional network, ResNet 18, trained on ImageNet for image similarity. ImageNet has 1000 categories, so this pre-trained network must learn relevant features to distinguish items from each other. From this network, we can extract the last embedding layer before the final softmax prediction. This leaves us with a 1000 dimensional vector for a given input image. These vectors can then be compared with L2 distance to get a metric for image similarity. Learn more about image search here. A dictionay keyed by objectID and with values of the vectors is serailized and stored to the pickle file,
-
- This notebook demonstrates the nearest neighbors algorithm working for the image feature vectors that were computed and saved in the
features_dict.picklefile. The images printed out to the screen should have objects that are visually similar.
- This notebook demonstrates the nearest neighbors algorithm working for the image feature vectors that were computed and saved in the
Sanity Check: After running all of the notebooks, your data directory should be set up with the following content:
imagesfolder with<objectid>.jpgnames: this contains to images for the objects in the dataset<filename>.csvfile: this contains the information for the objects.If using a
<filename>other thanmetdata, make sure to change"data/metdata.csv"to"data/<filename>.csv"in data_handler.py.features_dict.picklefile: this pickle file a serialized dictionary that holds a dictionary with key=objectid, value=1000 feature vector.
With the data processing complete, we can package our image search algorithm and output into a Flask server.
For this project, we only need a single endpoint--one which can accept a base64 encoded image and return the art recommendations is a clean JSON format. This section explains how to get the server running locally on port 5000, but these configurations can always be changed based on preferences. We will use port 5000 throughout this guide. The code for the server is located in server.py. See the code for more details about how it works. It is essentially a way of packaging the results shown in FindNearestNeighbors.ipynb into a server with textual content as well. Also, it is currently run with debug set to True (debug=True), but that can be changed to False if looking for faster performance.
After running the server locally (not in Docker), it will be located at at http://ipaddress:<port>/endpoint, where <port> is 5000 with the code included in this repository. We use a standard HTTP POST request in our Flask server.
-
Run the server.
# make sure to be using the virtual environment # start the Flask server cd main python server.py -
Request format design choices. Learn more about Flask POST requests on this page.
# input format image=<base64_encoded_image_as_string> # output format data = { "img_str": base64 string a combined image of the similar items (left to right will be most to least similar), "ordering": [objectid1, objectid2, etc.], # this output is organized in this way to be easily formatted with Unity, which is described in the Unity code "items_info": [ { "objectid": objectid1, "information": [ {"title": title information, "description": des 1}, {"title": title information, "description": des 2} ], }, { "objectid": objectid2, "information": [ {"title": title information, "description": des 1}, {"title": title information, "description": des 2} ], } ] }
This section outlines how to get the endpoint running in an Azure VM such that is is accessible on a Public IP address. This allows the HoloLens to have access to it.
-
Create an Azure VM
Follow this guide to create a virtual machine in Azure. After successfully completing the steps in this document, you logged into the Azure portal. On the left pane, click "Virtual Machines" and navigate to the VM that you created. Click on it, and you should be in the "Overview" pane. In this section, you will see a "Public IP address" field with a value. Take note of this because you will use it to
sshinto the VM to set up the server. To find your computer's IP address, you can go to go http://www.whatsmyip.org/. See this page for more details regarding white -
Access the VM and Set Up Server
Remote log into the VM by using a terminal and running the command
ssh "Public IP address", where "Public IP address" is the address noted in the previous section. If presented with an error, you most likely need towhitelistyour IP address. In the same menu as where the "Overpane" was located, click and open the "Networking" pane. Here you can add an inbound port rule or modify an existing one to allow your IP address to be allowed to connect to the server's "Public IP address". See this page for more details regarding whitelisting your IP address. Finally, you'll want to open port 5000 to be accessible; this is the port the web server will be using. You open this port to any inbound IP address, or you could specify the IP addresses you want to give access to. In this more secure approach, you will have to whitelist the IP addresses of your laptop running Unity, the HoloLens emulator, and the HoloLens device itself (depending on what you are accessing the endpoint from at the time).After remote logging into the VM (using
ssh "Public IP address"and following the username & password prompts), you need to get the repository and data one the machine. This can be done in two ways: (1) clone the repo and run the data preprocessing scripts there, or (2) remote file copy the repository (with thedatafolder ready to go) to the machine. This can be done with scp on Linux/Unix. Option (2) may take some time depending on how many images are int thedata/imagesfolder. -
Create Docker Image and Start a Container
Now we are ready to use Docker to deploy our Python server on the Azure VM's public IP address at port 5000. To complete the next steps, you will have to have Docker installed on the VM. Follow these steps to install Docker if needed. Finally, keep in mind that all the following commands are run on in the Azure VM terminal. Also, you may need to run
sudoin front of the docker commands depending on user privileges.# navigate the the repository cd path/to/the/repository # navigate the directory with the Dockerfile cd main # build the docker image in the directory docker build -t arart . # run the docker image in a container # this will start the server as specified in the Dockerfile docker run -d -p 5000:5000 -v $(pwd)/data:/main/data arart# SERVER WILL BE RUNNING AT THIS POINT. THE FOLLOWING IS ADDITIONAL INFO: # ---------------------------------------------------------------------- # (additional info) enter the container without starting the server right away # docker run -it --entrypoint /bin/bash -p 5000:5000 -v $(pwd)/data:/main/data arart # (additional info) enter a docker container that is already running # docker ps # make note of the container id # docker exec -it <container id> bash # notes about the docker flags: # -d runs it in the background # -p binds a host port : to the docker port # -v mounts a host directory : to the docker container directory
Sanity Check: After completing all of these steps and running the server, user a browser to navigate to http://"Public IP address":5000. "hello ar art app" should be written to the screen.
Please note that the parameters used in this tutorial are not strict and can be modified depending on user preferences. For example, a port other than 5000 could be used by modifying the -p flag parameter.
Here we explain the Microsoft Hololens application. We link some important tutorials and then explain where we deviate to create a custom experience with our custom HTTP endpoint. We use Unity for development, test with the Hololens emulator, and finally deploy to the Hololens device itself for demo purposes.
-
Set Up Unity to work with the HoloLens
Getting Unity set up for the Hololens is beyond the scope of this project, so we make references to Microsoft tutorials to get everything set up. Follow this tutorial to get set up with Unity and Hololens development. The start code from this tutorial is used heavily to get this project up and running--hence the name of the Unity application folder that remains in our repo: MR_ComputerVision. However, if your Unity development environment is already configured for Hololens and you understand how Unity app development works, feel free to load the code directory from the MR_ComputerVision without following the referenced tutorial. Once comfortable enough with developing with Unity for the HoloLens, go ahead and load the Unity project from MR_ComputerVision.
-
Some Notes about the Program
This C# scripts are documented well, but here we make a few notes here about things that need (or might need to be) changed before running the application.
-
Changing Custom API Endpoint in the Code
- The endpoint address will have to be changed to
http://"Public IP address":5000/endpointafter getting the server up and running in an Azure VM.
- The endpoint address will have to be changed to
-
Deciding Art Piece Information to Display
-
In the Unity code, we have a list specifying what type of information to display on the AR information UI panels. This list describes the fields that are available in the .csv that was created as part of the data formatting.
-
Here is an example of topics to display for a given art piece:
- Title - department - Culture - artistRole - objectEndDate - medium - creditLine - geographyType - classificationFrom these topics listed above, only the ones that are included in the POST response are written in the Unity world. This is important because sometimes not all entries are available in the .csv file. However, the assumption is made that the Title field is always present because every object in the dataset should have a name. You may need to edit the code if this is not the case for your data.
I've noticed a few times that the app has crashed for a key not found error. I believe I had these errors checked for in my code, but you may wish to verify and/or take a look if you experience problems with this.
-
-
We encourage you to get this code up and running with your own data. Applications for image similarity recommendations go beyond art. You could find related products, the celebraties people look closest, vacation destinations you may like, etc.
- Microsoft Academy Tutorials: https://docs.microsoft.com/en-us/windows/mixed-reality/academy
- Scaling Unity UI for HoloLens:
- Testing with the HoloLens Emulator: https://docs.microsoft.com/en-us/windows/mixed-reality/using-the-hololens-emulator
The steps to using AI locally on the HoloLens are not described in this repo. However, below are some resources and tips for how this would be accomplished for this project.
-
PyTorch -> ONNX Format
- The best way to run the inference for images locally would be to use the ONNX format as an intermediate format which can be ported to C# for use with Unity. We use PyTorch to create the feature dictionary, which was done in notebook 2_CreateFeaturesDictionary.ipynb, so we export the same model (contained in feature_extractor.py) to ONNX format. We can perform this operation with export_to_onnx.py. It creates an ONNX file called
"onnx_model_name.onnx", although this name can be modified. Note that image normalization will have to be done on the C# side if using the feature_extractor.py as is.
- The best way to run the inference for images locally would be to use the ONNX format as an intermediate format which can be ported to C# for use with Unity. We use PyTorch to create the feature dictionary, which was done in notebook 2_CreateFeaturesDictionary.ipynb, so we export the same model (contained in feature_extractor.py) to ONNX format. We can perform this operation with export_to_onnx.py. It creates an ONNX file called
-
Microsoft's Custom Vision API -> ONNX Format -> C# in Unity: https://meulta.com/en/2018/05/18/experimenting-with-windows-machine-learning-and-mixed-reality/
- This guide is a valuable resource that explains the pipeline from going from an ONNX model to using it in a Unity application for the HoloLens (or Mixed Reality device, in the article's case). We use a PyTorch model that can be exported to ONNX, so the steps after having the model in the ONNX format should work. However, image normalization will need to be taken into account.
-
General Structure:
- To make these steps work with the structure of this project, you'll have to edit the server to accept the vector computed from the ResNet 18 model. Essentially, by running the network on the HoloLens itself you eliminate the need for the server to run inference. This may save memory resources. The only the reason the server is still needed to run online is to hold the large number of images. However, you could opt to run the server on the HoloLens itself with all the images and the CSV stored locally. This would make an application that is entirely offline, but it may take a lot of space on the HoloLens.
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.
When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.