Merged dataset training

[lxtGH]

Pose track datasets have 15 keypoints while COCO has 17, how to merge the dataset and train Jointly?

[Guanghan]

Use posetrack format as the standard. What I did is mapping COCO to posetrack, which can be found in the following routine: change_order_COCO_to_PoseTrack.
(https://github.com/Guanghan/lighttrack/blob/master/HPE/AllJoints_COCO.py)

[lxtGH]

Thanks for your reply!

[lxtGH]

@Guanghan Hi! The training step for your model on pose tracking dataset is :

1. First train the COCO.
2. Then train on COCO and Pose Track Jointly
3. Finetune on Pose Track
   or somethings else ?