Protocols & Data Preprocessing on MPI-INF-3DHP

twehrbein / Probabilistic-Monocular-3D-Human-Pose-Estimation-with-Normalizing-Flows

This is the official implementation of the ICCV 2021 Paper "Probabilistic Monocular 3D Human Pose Estimation with Normalizing Flows" by Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn and Bastian Wandt.

MIT License

70 stars 6 forks source link

Protocols & Data Preprocessing on MPI-INF-3DHP #7

Closed GloryyrolG closed 1 year ago

GloryyrolG commented 1 year ago

Hi @twehrbein , @marco-rudolph , @bastianwandt ,

It is good to see the method achieves a good generalization of 86.9 on MPI-INF-3DHP. It would be better if you could share some details on how to reproduce the result, \eg,

if similar data preprocessing (-mean/std) as Human3.6M is adopted for 3DHP,
if so, if mean and std used are re-computed on 3DHP,
and if 3D output poses are PA-aligned (to remove scale and rotation) before computing errors.

Any help would be greatly appreciated.:)

Thanks & regards,

GloryyrolG commented 1 year ago

& if I were not wrong, it seems usually we use original 2D as input instead of cropped 2D coordinates, right?

https://github.com/twehrbein/Probabilistic-Monocular-3D-Human-Pose-Estimation-with-Normalizing-Flows/blob/ad2fdf21da1dfb689a5c2a531ebda6e23d95ccc0/data/data_h36m.py#L93

twehrbein commented 1 year ago

Hi @GloryyrolG !

We apply exactly the same data preprocessing for H36M and 3DHP. Note that we do not use statistics (i.e. mean and std) of the datasets, but simply center each pose to its mean and divide by its standard deviation. See code here. The predicted poses are not PA-aligned when computing PCK.

if I were not wrong, it seems usually we use original 2D as input instead of cropped 2D coordinates, right?

We use normalized 2D poses as input. Due to the normalization (see above), it does not matter if the 2D poses are in full or cropped image coordinates before normalization. "p2d_hrnet_unnorm" is saved only for visualization purposes.