search

bastianwandt / RepNet

This is the original RepNet implementation
86 stars 17 forks source link

Occlusions #8

Open Esbenthorius opened 4 years ago

Esbenthorius commented 4 years ago

Hi Bastian Thanks for the source code to this model, it is really well engineered. In your paper it is stated that the generator network can hallucinate points if there are not detected by the 2d pose estimation and that the spine points from the stacked hourglass model(SH) are not detected and therefore set to zero during training. However in the SH-data you provide for validation the spine point is located and if I set that to [0,0] i get really bad results. Am i doing something wrong? i really like the fact that this model should be able to handle occlusions

RobSalaets commented 4 years ago

I am struggling with the same issue

lvandoit commented 4 years ago

I noticed that in the discriminator these was a wassertein loss. But in the code it was defined as 'mean(y_true * y_pred)'. So there had any precess for it. @RobSalaets @Esbenthorius