search

kristijanbartol / linear-3d-humans

"Linear Regression vs. Deep Learning". The source code for a simple but effective baseline method for human body measurement estimation using only height and weight information about the person.
Other
6 stars 2 forks source link

Inconsistent height estimates #1

Open jutanke opened 1 year ago

jutanke commented 1 year ago

Thanks for making your project available and so easy to use!

I played a bit around with your demo and found some inconsistencies with the estimations: I used your demo inputs: male -> 1.72 65 and got as SMPL shape estimation:

[ 1.05071094  0.72027541  0.2989638   0.05701623  0.07447251 -0.09911867
  0.07585485 -0.05481659 -0.00452416  0.11355854]

However, when I use a larger person: male -> 1.83 74 I get the following SMPL shape estimation:

[-0.2499936   1.10689084  0.38673149  0.10451252  0.08272817 -0.12257114
  0.04098995 -0.08785101 -0.03187838  0.10152087]

As the first principle component (kind of) represents body height it seems that the second person estimation is much smaller then the first one, even though the second person is more than 10cm taller:

frame00000 In the above image the blue skeleton represents the 1.83m tall person while the red skeleton represents the 1.72m tall person (same pose for the two given shape parameters).

Interestingly, if one uses: male -> 1.80 50 one gets the following shape:

[ 0.52775807  2.48606748  0.51797164  0.17677026  0.05246255 -0.41012379
  0.03961402 -0.1975367  -0.00779281  0.17064792]

while when selecting male -> 1.80 100 one gets:

[-0.45392014 -0.95924672  0.15785673 -0.02097742  0.11747412  0.27203582
  0.0648735   0.07791602 -0.04637479  0.01784527]

which is weird considering that both persons have the same body height.

Is there anything I am doing wrong? I simply print shape_params (https://github.com/kristijanbartol/linear-3d-humans/blob/c72a2c1abb756054aa0e2c018078e61fad62e171/demo.py#L26)

kristijanbartol commented 1 year ago

Hi, thanks for your interest! I apologize for the inconsistencies. It is not impossible that there is a mistake in the public version of the demo. Unfortunately, I will be able to check it only after the New Year. In the meantime, you can check whether the estimated body measurements are also inconsistent. If yes, then surely there is a mistake. Sorry about that, I did additional experiments with the code after we published the paper, so I have to clean it in any case!