Bug report on densepose_cocoeval.py

[penincillin]

Hi, I am currently reading the source code of densepose, and I think there might be a bug in detectron/datasets/densepose_cocoeval.py. Don't be nervous, this bug won't affect the result of the code, it will just confuse people like me who try to read your code :)
From line 378 to line 397 I think some 'x' and 'y' are messed up and the code with correct use of 'x' and 'y' should be :
Line 378~379:

dy = dt['bbox'][3] 
dx = dt['bbox'][2]

Line 382~383

py = ( dp_y + g_[1] - dt['bbox'][1]).astype(np.int)
px = ( dp_x + g_[0] - dt['bbox'][0]).astype(np.int)

Line 395~397

ipoints = dt['uv'][0, py, px]
upoints = dt['uv'][1, py, px]/255. # convert from uint8 by /255.
vpoints = dt['uv'][2, py, px]/255.

I would appreciate if you reply me.
Thanks !

[vkhalidov]

@penincillin Indeed, the names x and y are partly interchanged here. Bounding box dimensions dx and dy computed on lines 378-379 and point coordinates px and py computed on lines 382-383 refer in fact to y and x coordinates respectively. At the same time, dp_x and dp_y variables refer to x and y respectively. This creates syntactic confusion, but the code is semantically correct.

[penincillin]

Thanks for your reply !
I got it.

[Johnqczhang]

@vkhalidov @penincillin I fixed this confusion and simplified the code in #203, which makes the code understood more easily and has no effect on the results.