Training Data of pretrained models

[lorenz-lb]

Hello,
i am researching with ViTPose (https://mmpose.readthedocs.io/en/dev-1.x/model_zoo/body_2d_keypoint.html#topdown-heatmap-vitpose-on-coco).

If I want to train a ViTPose model on lets say MS COCO, do I need to use the pretrained checkpoint?
The official paper "ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation" states that pretraining on only MS COCO achieved an AP of about 75 on MS COCO val (seen in Table 3).

Furthermore If i wish to use the pretrained model provided by mmpretrain, how do I know which dataset was used to train this pretrained checkpoint? Was it only trained on ImageNet-1k or did the training data include COCO data (as described in the paper also)?

If I want to train on (a modified version of) COCO, I can't allow COCO data in the pretrained data.

Thanks for your help! And thank you for this wonderful community and repo 😄