The predicted skeleton is slightly deviated from the body about rmpe

[Fang-Haoshu]

Hi! Thanks for reaching out. The problem may be that we crop the single person image first and re-project the pose results to the original space at last, and this procedure might introduce some small bias. You can add 1 pixel to the final pose results.

[RuWang15]

Thanks for your reply! Now I want to implement the demo on VS2015, but your version of caffe doesn't support Win well. Installing caffe on VS is really frustrating and I'm still trying... Do you have any plan to develop a windows version or can you give me some tips on installing it? Thank you!

[Fang-Haoshu]

Hi,
Sorry, I haven't implemented it on Windows and have no plan to do so. I am not familiar with Caffe on Win...

Hope you can figure it out soon!

[RuWang15]

I add some of your layers in caffe-win, and try as I might, I cannot run VGG_SSD model successfully. But fortunately I can run shg+sstn model with a bounding box detected by myself...
Moreover, I found that the res param in function transformBoxInvert() should be a little bit smaller than 64, so that the points can be zoomed properly...Is there an exact number that can produce precise result?

[Fang-Haoshu]

Congratulations!
Sorry, I didn't experience with that...