Incorrect SMPL ply from h4d_fitting.py from kinect v2 depth_pcl

[remmel]

Hello!

Using my own depth_pcl ply, I fail to create the SMPL via h4d_fitting.py, the ply produced is an ugly deformed human! (and hopefully I did not resemble to it :)
The ply I created comes from the kinect v2 (512x424) and was filtered semi-manually. The kinect is looking forward (not tilted), landscape, positionated at the level of my chest (90cm). The depth is noisy as it comes from a Kinect and I'm doing simple poses.

I ran PYTHONPATH=. python h4d_fitting.py --pcl_type depth_pcl --seq_name 00002_20240109-1553-kinect2-py_0

As you can see my poses are simple (here the point cloud is not manually filted yet)

But that deformed human is created (for the ply I rotate in the gif - 0002.ply )

See log.txt attached, first and last step:

Step 1 (1/500) | ETA: 0:13:09 | batch_time: 1.58 | lr:0.030000 | point2surface: 0.015792 | betas: 0.0000 | pose_pr: 24.7866
Step 500 (500/500) | ETA: 0:00:01 | batch_time: 0.76 | lr:0.009830 | point2surface: 0.000833 | betas: 67.3510 | pose_pr: 909.3285

In that link, RGB and D (16b png), depth_pcl colored plys and h4d_fitting plys : https://drive.google.com/drive/folders/1P61uttP9okI0XSD1yWEpG_Uz7gf-9YWN?usp=drive_link

Do you have some tips? wrong origin or orientation? too noisy point cloud?

Note that

• I successfully created the SMPL for 02474_longshort_ROM_lower_258 python h4d_fitting.py --pcl_type depth_pcl --seq_name 02474_longshort_ROM_lower_258.
• I also unsuccessfully tried with images from my Kinect DK.

Thanks for your help!

Rémy