During the training of PointNet all point clouds are normalized into a unit sphere.
Is this step necessary when the scale of the point cloud is significant for the task at hand?
For example, if we want to regress the volume of the point cloud, the its scale matters. Since the normalization is not aware of the other point clouds, this can confuse the network. For example, assume we have two 2D point clouds (for simplicity). Point cloud A has points {(1, 1), (-1, -1), (1, -1), (-1, 1)} and point cloud B has points {(2, 2), (-2, -2), (2, -2), (-2, 2)}. After normalization, both point clouds will become the same point cloud (although they were different, with different initial volumes).
During the training of PointNet all point clouds are normalized into a unit sphere.
Is this step necessary when the scale of the point cloud is significant for the task at hand?
For example, if we want to regress the volume of the point cloud, the its scale matters. Since the normalization is not aware of the other point clouds, this can confuse the network. For example, assume we have two 2D point clouds (for simplicity). Point cloud A has points {(1, 1), (-1, -1), (1, -1), (-1, 1)} and point cloud B has points {(2, 2), (-2, -2), (2, -2), (-2, 2)}. After normalization, both point clouds will become the same point cloud (although they were different, with different initial volumes).