We’ve got some exciting Climbuddy news!

As mentioned in the last update,

Besides many improvements described in the nerd stuff section below, we’re a week or two away from finally launching a testing instance for Crimp, a climbing gym near Heidelberg, Germany – the code that produces the models is in a relatively mature state, with most of the work being spent on UI/backend.

we did the classic forget to multiply the estimate by two (there were many more things we wanted to pack into the first stable release), resulting in the instance being up now – head over to

A poster for the Crimp testing instance.
A poster of the Crimp testing instance (available at https://climbuddy.com).

to check it out and let us know what you think!

Nerd Stuff

We’ve done some experiments with YOLOv8 as an alternative segmentation network to Detectron2, and it looks like we will be using it instead. It’s much more ‘batteries-included’, which fits what we’re trying to do more than having to fiddle with Detectron2 parameters.

We’ve also tuned parameters for determining objects of interest (that are later segmented). We use adaptive thresholding (as described in OpenCV’s documentation), reimplemented to work for meshes – this works much better than global thresholding, since two different holds can be seen by a wildly different number of images.

Thresholding comparison.
Mesh thresholding with old parameters (up) and new parameters (down).

Team Climbuddy