Visualizing on custom dataset

[ut-austin-rpl]

Visualizing on custom dataset

[Steve-Tod]

I think it should be in [0, 0.3] * 3 coordinates.

[ishikaalunawat]

I think it should be in [0, 0.3] * 3 coordinates.

You mean the scale of range of the grasp query or the scene?

[Steve-Tod]

I think it should be in [0, 0.3] * 3 coordinates.

You mean the scale of range of the grasp query or the scene?

The scale of the scene.

[ishikaalunawat]

I think it should be in [0, 0.3] * 3 coordinates.

You mean the scale of range of the grasp query or the scene?

The scale of the scene.

Right.
So basically I made the TSDF (voxels) such that each voxel has a size 0.0075 cm and so the TSDF has length of 30cm in a (40,40,40) grid.

But in Trimesh my scene seems to be shifted from the origin, so how did you scale it in your case? Is this just a trimesh thing?

[Steve-Tod]

I think it should be in [0, 0.3] * 3 coordinates.

You mean the scale of range of the grasp query or the scene?

The scale of the scene.

Right.

So basically I made the TSDF (voxels) such that each voxel has a size 0.0075 cm and so the TSDF has length of 30cm in a (40,40,40) grid.

But in Trimesh my scene seems to be shifted from the origin, so how did you scale it in your case? Is this just a trimesh thing?

Does it look like a uniform shift? Because the prediction space of implicit functions is [-0.5, 0.5], I scale and shift the output into [0, 0.3] in the visualization function.

[ishikaalunawat]

Thanks for the help! To sum up the ranges:

input query points : [0,0.3]
output of implicit function : [-0.5,0.5] which is then scaled & shifted to get it in [0,0.3]

Is this correct?

[Steve-Tod]

Thanks for the help! To sum up the ranges:

input query points : [0,0.3]
output of implicit function : [-0.5,0.5] which is then scaled & shifted to get it in [0,0.3]

Is this correct?

Yes!

[ishikaalunawat]

I needed some help regarding sampling of query points. I've used your sample_grasp_point() function. But how did you choose the threshold value for normal[2] > -0.1 for selecting a surface point? The problem I faced is while sampling points, points inside the object are also present.
For reference,

Image of the point cloud	Normals of point clound

Note: I've used pcd.estimate_normals() to get the normals.

[Steve-Tod]

Hi, it's OK to sample points inside the object, because in that case, the gripper would collide with the object before getting to the target position, and the current data generation script would detect that collision and mark the grasp as a failure.