zhangboshen / densenet-human-pose-estimation Goto Github PK

Hi boshen:
i fork your project and run on my own datasets. i resize the train images to [200,200], but the number of feature points (label) in my trainsets is 16 points. I modified your code in order to meet my demand: all 22 in your code are changed to 32. I have two questions to be solved.

1. How long time do you spend to train the model? I run on two 1080Ti GPUs(10G memory each),i have trained 50000 steps, but the loss is still10000, which I think is a little high. I want to know the time or steps you spent on training.
2. The label. I see your label is XXX.mat with [num_img, num_points, 2].
   Assume that my label file is Label.mat. so the size(Label) = [6131,16,2]
   and the Label[ : , : , 1] represents the x-coordinate of the featue points ( row in image).
   Label[ : , : , 2] represents the y-coordinate of the featue points ( col in image ).
   Is the format of my Label the same as yours?

Thank you very much.