Train on custom 2D pose dataset

[alexanderfrey]

Hi and thanks for great work !

I would like to finetune the model on some custom 2D pose dataset. From what I read you trained on 3D and 2D pose datasets so that should be possible, right ?

Best
Alexander

[Arthur151]

Yes, of course. You can refer to this issue to create your own dataset.

[alexanderfrey]

Thanks for your answer. That particular issue is about retraining on the model output. My question is more related to how the data in the img_info dictionary is generated and what is the format ?

I have right now 2D pose data in a csv file with the format nose,182,416,img001.jpg,768,800 etc.

How do I go from that data to the img_info dict ?

Any help is very much appreciated.

Best,
Alexander

[Arthur151]

Crowdpose is a perfect example for your need. It just loads the 2D pose data. You need to replace this line to the 2D pose of your data. Please be sure to set the correct keypoint ordering at this line. To achieve this, you can define a dict in constants.py like this. Then you can replace constants.Crowdpose_14 to the keypoint order dict you defined.
self.kp2d_mapper = constants.joint_mapping(constants.KEYPOINT_ORDER,constants.SMPL_ALL_54)