frankier / skelshop

📺 📰 🧑‍💼 Toolkit for skeleton & face analysis of talking heads (e.g. news) videos 🧑‍💼 📰 📺
https://frankier.github.io/skelshop
MIT License
5 stars 1 forks source link

What are pose-matcher-configs? #28

Closed cstenkamp closed 3 years ago

cstenkamp commented 3 years ago

While track-confs are understandable as one of click.Choice(TRACK_CONFS.keys()) I don't understand what a pose-matcher-config is even? What is supposed to be stated there as Text?

frankier commented 3 years ago

This is for the skeleton based reidentification. The models from lighttrack are used for this https://github.com/Guanghan/lighttrack

This is only needed if you use the lighttrackish tracker, but this is currently the best I think.

Everything is handled by default for the current best settings by the Snakemake configuration: https://github.com/frankier/skelshop/blob/master/workflow/Snakefile https://github.com/frankier/skelshop/blob/master/workflow/rules/skels.smk

Documenting usage via Snakemake is on the TODO https://github.com/frankier/skelshop/issues/21

frankier commented 3 years ago

In short, if you follow the installation instructions or use the Dockerfile, you should already have it at work//gcn_config.yaml

frankier commented 3 years ago

Here's where the magic happens https://github.com/frankier/skelshop/blob/dfff53eeaa7ec47ff064b6a793a29f0e8d85e79e/skelshop/pipeline.py#L81

But this whole part is one of the ugliest bits of this whole thing because we have to pull in a bunch of horrible stuff to get the pretrained lighttrack model to work. It would be nice to clean this up at some point, but perhaps it's less important if we switch to a pose estimator which already does tracking.

frankier commented 3 years ago

Sorry that should have been opt_lighttrack is the best

frankier commented 3 years ago

Closing as answered. Feel free to reopen.