the result

[vicdu]

the result

[mkocabas]

Hi @vicdu,
The result of the original code is on the MPII test set. In general, test accuracy is ~2 point higher than validation accuracy. @bearpaw's implementation give better accuracy than original one. You can take a look at these discussions:
princeton-vl/pose-hg-demo#1
princeton-vl/pose-hg-demo#6
#15

[heroxx2011]

Hi vidcu, I have trained the model and the result only 84.89. I think maybe the parameters problem. Welcome to discuss.

[vicdu]

@heroxx2011 hi, Have you read the code in torch7 ? I found something like '+offset' in dataload part. And, do you think it will be beneficial to expand the mpi-dataset?

[bearpaw]

Thanks! @mkocabas the explanation also matches my own practice. Also, this repo is a reimplimentation of the original hourglass with PyTorch, so we cannot expect the exactly same results.