B-SOiD v1.3 (hello, python!)

We are proud to release a major update for the B-SOiD Python suite (bsoid_umap). The MATLAB package is not changing. In both cases, the user experience is no different. If, however, you are new to python package, you can watch the video in demo/bsoid_py_tutorial_v2.mp4.

By implementing UMAP with HDBSCAN, B-SOiD in Python can now handle high-dimension clustering with an essentially infinite number of features (we’ve tried up to 784 features = 28 body parts).

Moreover, this update renders the code truly agnostic - not just to the groups extracted, but to the input data itself. Although B-SOiD has always been capable of taking in any set of spatiotemporal data, it is now optimized to handle any set of estimated positions, from any camera angle, and of any model system.

As always, once trained, B-SOiD can be applied and generalized across datasets, processing hundreds of thousands of frames a minute - and do so with the same temporal resolution as the camera being used.