Questions regarding inference on the custom data

[yohanshin]

Hi, thank you for your great work! I would like to run your model to triangulate the data that we've collected.

Since we are using detection models with COCO configuration, I am not sure how can I run your model with COCO.

Do we have to retrain the model on COCO configuration?

In infer_pose.py, it seems like I need h36m_custom.pt while I can only download h36m_knwon.pt and h36m_est.pt from your shared link. Or can I use h36m_known.pt if I already have the camera calibration parameters?

Thank you for your answer in advance!

[kristijanbartol]

Hi @yohanshin , I apologize for my late response! You should be able to use h36m_known.pt. The h36m_custom.pt comes in in case you have your "custom" parameters, i.e., estimated by a different method.

For using COCO body models, you should retrain the model on such a body model because the joint semantics is a bit different than the H36M body model. The problem is, I think that H36M dataset contains only H36M ground truth annotations? I would have to check, but I think that CMU might contain COCO annotations? Then you could train the model this data and apply it in inference to your collected data.