demo

Running the Standalone Demo

The easiest way to test this project is by running the standalone demo. Download the trained model from Github and put it into this directory.

Then run standalone_inference_over_image.py from within this directory:

$ python standalone_inference_over_image.py --detection_inference_graph 2019-04-24_faster-rcnn_inception-resnet-v2.pb --input_image IMSLP454437-PMLP738602-Il_tempio_d_amore_Scene2-0002.jpg --output_result output_detections.json

This will create an annotated image, as well as a json-file with the output detections.

Running Inference as a Service

The second way to run inference is by firing up an inference server that exposes a REST API for easy consumption.

To do so, start the server script with hug -p=8080 -f=inference_server.py (see also run_server.bat) or create and run a docker container with the following steps:

# Build the image with the latest model
$ docker build -t measure_detector .

# Run in container (change port to `XXXX:8080` if needed):
$ docker run -p 8080:8080 measure_detector

⚠️ WARNING: The server should only be used for testing, not deployment!

Talking with the REST API

We offer two example Python 3 scripts. Make sure that you have all the requirements:

$ pip install -r requirements.txt

Single Image file

test_rest_api.py show a very basic Python script for a single image.

$ python test_rest_api.py IMSLP454435-PMLP738602-Il_tempio_d_amore_Sinfonia-0011.jpg

Generating an MEI file from a folder with score images

folder_to_mei.py detects the measures in all images of a given folder and generates an MEI file for further processing. --make-images is optional and generates images with bounding box overlays in a subfolder.

$ python folder_to_mei.py --make-images IMSLP108695