Why is there a preprocess that masks out voxels whose distance to the nearest surface is smaller than the finger depth?

[Fanqyu]

Hi there!
In the paper, it said that "mask out voxels whose distance to the nearest surface is smaller than the finger depth as defined in Figure 2(c)". I feel confused about this process as your code implementation (/vgn/src/vgn/detection.py) obviously masks out those voxels either are outside the surface (tsdf < 0.5) or are just on the surface (tsdf = 0.5).

    # mask out voxels too far away from the surface
    outside_voxels = tsdf_vol > 0.5 
    inside_voxels = np.logical_and(1e-3 < tsdf_vol, tsdf_vol < 0.5)
    valid_voxels = ndimage.morphology.binary_dilation(
        outside_voxels, iterations=2, mask=np.logical_not(inside_voxels)
    )
    qual_vol[valid_voxels == False] = 0.0

It seems that only the grasps whose positions inside the object are possibly executed then. From my understanding, masking out voxels too far away from the surface is natural and necessary. However, what is described in the paper is confusing as illustrated by the following image:

The grasp described in the image is obviously a valid grasp.
From my understanding, I guess utilizing grasps from voxels inside the object is to make grasps approach the object surface very closely so as to make grasps tight and robust.
But in your video, the generating grasps which are visualized are not as close to the surface as I think. Did I miss something important? Could you give me some instructions?