Jennifer J. Sun1, Jiaping Zhao2, Liang-Chieh Chen2, Florian Schroff2, Hartwig Adam2, Ting Liu2
Caltech1 Google Research2
These models are trained with the publicly available data processing script and code.
Keypoint detector = CPN Keypoints finetuned on Human3.6M (download instructions in the data processing script)
Note that since the model is trained with CPN keypoints, it will not perform as well using other keypoint detectors.
If you would like to contribute a trained model to this page, please contact us: jjsun (at) caltech.edu.
Pre-trained model Checkpoints for POEM:
Pr-VIPE, 16 dim embeddings, trained on Human3.6M: [Google Drive Link]
L2-VIPE, 32 dim embeddings, trained on Human3.6M, with random dropout (occlusion-robust): [Google Drive Link]
See the "all_flags.train.json" for more details on the hyperparameter used to train each model.