Partial pose on full_human pretrained model

Image Analysis
#1

Hello,

Does the full_human pretrained model from DeeperCut must require fully visible pose in the video? Is it possible to modify it somehow so that it will work to test on partial pose?

I wanted to test my infant data on pretrained human adult network. However, the results are not very good (because the infant videos are partially pose) so I need your help to clarify.

Thank you,
Mai

#2

Hello @buim! No, it is not a requirement, although the pretrained model is known to work better for videos of about 300 pixels in height. So if your videos are larger, resizing may help :slight_smile:

#3

Hello @jeylau,

What I mean is that my infant data is only showing upper body of the baby. When I used pretrained model (full_human model) on the infant data, the results are not so good. The labeled video does not label accurate keypoints or not label any keypoint at all.

#4

I suppose the model was trained either only or mostly on videos with full pose visible so you might need to retrain a bit on your data.