Why does GraspNet inference require reversed points as input?

[Terence233]

Hi there,

Thx for the amazing work! I have a question regarding the usage of GraspNet.

In your implementation, I noticed that the inference step requires reversing the point cloud:

def compute_grasp_pose(self, full_pcd):
    points, _ = o3dp.pcd2array(full_pcd)
    grasp_pcd = copy.deepcopy(full_pcd)
    grasp_pcd.points = o3d.utility.Vector3dVector(-points)

    # generating grasp poses.
    gg = self.graspnet_baseline.inference(grasp_pcd)
    gg.translations = -gg.translations
    gg.rotation_matrices = -gg.rotation_matrices
    gg.translations = gg.translations + gg.rotation_matrices[:, :, 0] * self.config['refine_approach_dist']
    gg = self.graspnet_baseline.collision_detection(gg, points)

My questions are:

1. Why does graspnet_baseline.inference() here take the negated points as input?
2. Is this related to the oracle “ideal camera” setup?
3. Or is it due to a left-handed vs right-handed coordinate system issue from OpenGL rendering?

Any clarification would be really helpful. Thanks!