Question for changing df_voxel to .obj

[LV0913]

In test.py, save_vox_to_obj(model.voxel.data.cpu().numpy(), 0.5 if not use_df else 0.85, obj_name). Why th parameter 'th' is 0.85 instead of 0.01 for default of ios_th.
def save_vox_to_obj(voxel, th, save_path): if len(voxel.shape) == 5: voxel_iso = voxel[0, 0, :, :, :] verts, faces, normals, values = measure.marching_cubes_lewiner(voxel_iso, th, spacing=(1 / 128, 1 / 128, 1 / 128),) save_obj(verts - 0.5, faces, save_path)
In the above usage, why verts - 0.5?
Beacuse in test_shape.py,
v, f, n, _ = measure.marching_cubes_lewiner(output, opt.ios_th, spacing=(space, space, space)) save_obj(v, f, join(shape_dir, '%04d' % idx))
I wonder the difference for two cases. Thanks!

[ztzhang]

1. the thresholding level is different since the data stored is distance function, which is stored as the distance between the voxel grid's center to the nearest surface. We convert the distances into probabilities using a zero-mean gaussian. Therefore, 0.85 indicates a small region around the GT surface. We didn't really use the 0.01 as a threshold for df_voxels.
2. The shifting in vertices is just for the convenience of visualization. Shifting by 0.5 makes the mesh centered around zero. This should not affect the final output though; if you use mesh lab to visualize the output, the mesh is always zero-centered.

[LV0913]

Thanks for your reply. So in the train of the net, you convert the distances into probabilities using a zero-mean gaussian. That is why in test.py you use 0.85 for model.voxel.data.cpu().numpy() because it's probabilities. And when in test_shape.py, output is in distance function. So 0.012 is used in test_shape.sh.Is that right?
How can I Corresponde the th for distance function and probabilities?
I have figure out the second problem thanks for your help.