fc8 or fc8_softlabel?
vj-1988 opened this issue · comments
Dear @liuziwei7 : I am currently converting the demo.m from matlab to python since I don't have matlab. I am currently using the output from 'fc8' layer after forwarding the image. But output from fc8 are in negative float values. Can you confirm if this is as expected.
Hi @vj-1988
You can just use "fc8_landmarks" to get prediction of landmarks. "fc8" is the combination of landmarks and visibility.
And fc8_landmarks has been normalized to [-0.5, 0.5], so I think that is not as expected.
Hi @liuziwei7 : I have used fc8_landmarks to get the prediction of landmarks. But still i am getting negative values in the result.
Here is the result i got for one image on FLD_full model:
[[ 0.27906671 0.06566589 -0.00434459 -0.11829091 0.2226638 0.35514393
-0.02583538 0.11681057 0.00701905 0.09816497 0.19388133 0.01659511
0.22986662 0.01280492 0.30807292 -0.16883679]]
Any help will be appreciated. Thanks
Actually "fc8_landmark" represents centered and normalized fashion landmark coordinates.
Thus, the outputs of "fc8_landmark" are floats between [-0.5, 0.5], which could be either positive or negative.
Yes, the results you obtained are as expected.
You can transform these floats into absolute coordinates by executing the inline function "get_orig_coordinate" (Line 35 in file "./scripts/pipeline_forward.m").
I am getting the absolute coordinates, but those are not fitting for the input image. Are those coordinates obtained by that method corresponds to 224x224 resolution? Just clarify me on this.
Yes, the obtained absolute coordinates are relative to input image with size of 224*224.
Hi @liuziwei7 : Could you provide me the order of landmark points for full, upper and lower models? I am confused whether the first coordinate is for left arm or shoulder. I tried fitting with different combinations, but getting wrong landmarks. So just needed to verify whether i am fitting in wrong places? Thanks.
The order of fashion landmarks is given in the README file of "Fashion Landmark Detection Benchmark": http://mmlab.ie.cuhk.edu.hk/projects/DeepFashion/LandmarkDetection.html . It should be: "left collar [x, y]", "right collar [x, y]", "left sleeve [x, y]", "right sleeve [x, y]", "left hemline [x, y]", "right hemline [x, y]".