DEBobo is a library providing an easy-to-use evaluation code for object detection models. It's main motivation was to replace the part of pycocotools responsible for evaluation as it doesn't work well with custom datasets. Additionally, I found the workflow of pycocotools doesn't work well with high-level training libraries such as ignite.
Most end users should be able to get away with
pip install debobo
For development, install the library with symlinks and with the additional test requirements using
pip install -e .[test]
To run the tests, first download the test-data using ./fetch_testdata.sh.
The test is run via
pytest tests/
It compares the results obtained from debobo to the result obtained using pycocotools.
We provide ready-to-use metrics for ignite in debobo.adapters.ignite.
Feel free to request other adapters as an issue.
Also, check the tests for how to use the low-level routines.
Thanks to @martiningram for the header image.