Lours (the bear in French) is a library developped at XXII to help manipulating object detection datasets.

Powered by pandas under the hood, Lours will let you manipulate an object detection dataset as if it was a simple dataframe.

Cumbersome operations like conversion between dataset formats, splitting or merging are all accessible through a simple yet powerful fully documented API that will let you focus on actual research related tasks.

In addition to dataset manipulation, Lours offers a powerful evaluation feature which lets you visualize the influence of parameters across your dataset for a specific metric.

Documentation

See the full documentation, with API specifications and tutorials here : https://lours.readthedocs.io/latest/index.html

Code

Code is available on github. PRs and Issues are welcome !

Talk

To present our work, I have given a talk at the Computer Vision Paris meetup, at Theodo’s headquarters.

In addition to present Lours feature, this presentation featured a real world case study where engineers at XXII used Lours in their workflow to train a detection model, evaluate it and elaborate a set of camera and scene specifications for guaranteed quality of detection.

Thanks Urszula Czerwinska for the nice writeup, and thanks Theodo for hosting the event !

Slides