Computer vision based planogram compliance evaluation. Code for the master's thesis of Julius Laitala, University of Helsinki, 2021. The thesis is available at http://urn.fi/URN:NBN:fi:hulib-202106092585 .
Currently, the functions in cvpce are set up to run only on CUDA.
Therefore,
you'll need a NVidia card to run cvpce.
We suggest using Conda
to avoid CUDA installation pains.
To create a Conda environment for cvpce,
simply utilize the provided environment.yml:
conda env create -f environment.yml
conda activate cvpceWith the Conda environment set up and activated, cvpce can be installed with setuptools:
pip install .If you wish to tweak cvpce a bit,
the -e flag is your friend!
cvpce is a command line tool,
and a bunch of usage instructions can be accessed with the --help option.
Go ahead and
cvpce --helpafter installing to explore the available commands!
Pre-trained model weights are available in the releases.
The following public datasets were used for training and testing cvpce:
- GLN training and product proposal generation testing: SKU-110K of Goldman et al. (2019)
- DIHE training: the 2019 version of GP-180 (Tonioni et al. 2017)
- Classification and product detection testing: the Grocery Products dataset of George et al. (2014) with annotations from GP-180 (the 2017 version)
- Planogram compliance testing: planograms from GP-180 (2017 version) with fixes from this gist