Question about the documentation

[ldpbuaa]

This is an awesome project for pose estimation. Thanks for sharing the code! I would be appreciated if the author could append the documentation for data preparation, training, and testing. Is it possible to use the code to train LSTM pose machine on Penn Action dataset now?

[pasawaya]

Hi @ldpbuaa,

It's actually really straightforward to train the LSTM Pose Machine model on the Penn Action dataset. All you'd have to do is run python train.py --model=lpm --dataset=PennAction ... with your desired hyper parameters passed in as additional arguments (for a full list of supported arguments see the main function in train.py).

[ldpbuaa]

Thank you. I'm trying it. BTW, did you use the pretrained CPM model when train the LPM on the Penn Action dataset?

[pasawaya]

For pre-training, I followed the LSTM Pose Machines paper in training a two-stage single-image CPM model on a combined MPII and LSP dataset. The weights were then copied across all T stages of the sequential model since the paper also prescribes weight sharing.

[ldpbuaa]

The architecture of this project is really nice. I can train different models on different datasets with a tiny change of the code. But I noticed that there are some differences in the implementation of the test stage. The origin matlab code uses two deploy.prototxt files when testing, which processes the first frame separately in depoly1.prototxt as illustrated in the paper. Does this have an influence on the test result?

[sklf]

Hello，I am trying to train the LPM foolowing the instruction described in original paper. But I don't know how to pretrain the model on MPII and LSP dataset. Can you guys tell me how to do that? Thanks!

[YHDang]

hello, I have run the code successfully. But, when I calculate the PCK using the function get_preds(·), the results are 0. I wanna know if I process the coordinates obtained by get_preds(·). Looking forward to your reply, Thanks!