about the mesh vertices format(xyz or uvd) about arielai_youtube_3d_hands HOT 3 CLOSED

Hi,I am not the author of this paper,but I would like to share some my understanding.
Actually,Yes, you are right,The output of the network is generally speaking '(u,v,d)'.I wanna talk about the GT mesh generation step,e.g the “itterative fitting” method in the paper ，In fact this optimization based method did generate 3D hand mesh in camera C.S，But for better convergence for spiralNet ,the Arthur align these vertices to image C.S,which in your words '(u,v,d)'，
And this is why spiral-cov-net working because of this strict constraint of "point(x,y)-to-point(u,v,d)",so you can regard these approach as a "dense 3D hand keypoints estimation task".some of the other scholars work such as https://github.com/mks0601/I2L-MeshNet_RELEASE also did the same thing

As for normalization of mesh‘s gt,you can use following code:
`
train_db_np, val_db_np = np.array(data_splits['train'], dtype=np.float32), np.array(data_splits['val'], dtype=np.float32)

mean_points = np.mean(train_db_np, axis=0)
std_points = np.std(train_db_np, axis=0)

# Normalize vertices.
train_db_np = (train_db_np - mean_points) / std_points
val_db_np = (val_db_np - mean_points) / std_points

`

This is also @dkulon ‘s work

from arielai_youtube_3d_hands.

@dumyy,Actually the gt of mesh already in 3d format（Although we call it 'u v d'）,you could use meshlab to visualize these vertices,just like below:

(BTW the faces are from mano website)

from arielai_youtube_3d_hands.

@quyanqiu In fact, I have also another confusion. If I have got the mesh output, or just use the gt of mesh, how to project the points to 3d format ( in camera C.S. or just root-relative C.S.)?

from arielai_youtube_3d_hands.