measure-detection

A measure detection model for typeset scores, based on the tensorflow object_detection api. Inspired by the more general MeasureDetector for handwritten and typeset scores.

This repository uses the AudioLabs_v2 dataset you can get it here

training

note: please make sure you use tensorflow<2.9, on windows using the latest version of tensorflow will cause it to break JIT compilation

it is recommended to have make installed to build the targets, however you can always run the scripts manually

if you want to change the command the makefile uses to run python, you can change the PYTHON environment variable.

before training one should make sure that the tensorflow object_detection api is setup correctly

git clone https://github.com/tensorflow/models.git
cd models/research
protoc object_detection/protos/*.proto --python_out=.
cp object_detection/packages/tf2/setup.py .
python -m pip install .
cd ../..

quickly test the installation with

make test

then download a pretrained model from the tensorflow model zoo like ssd_resnet50_v1_fpn_640x640_tpu-8, unzip the contents and copy it into the pretrained directory

your directory structure should look like this

models/
    ...
pretrained/
    ssd_resnet50_v1_fpn_640x640_tpu-8/
        checkpoint/
        saved_model/
        pipeline.config
...

after that, configure the path to the dataset with the DATASET_DIR environment variable and run

make prepare-dataset

we are now ready to begin training

make train
make evaluate

after the training has finished, you can export it to a saved_model using

make freeze-pb

then perform a sanity check with

make inference-pb

and if you wish to you can convert it to a tensorflowjs model using

make convert-tfjs