Question about the position of the object center with respect to the global frame

[jjpatino-byte]

          Thanks for your question. The point cloud obtained from the sensor should be transformed to the global frame (where Z+ is the normal vector of the horizontal plane, e.g., the tabletop or the ground) since the output from our pre-trained network is constrained to a predefined view space in the global frame.

Originally posted by @psc0628 in #2 (comment)

Taking into account that the pointcloud should be transformed to the global frame. Does it matter where the object center (which is the same as the hemisphere center) is with respect to global frame?

[psc0628]

The object needs to be centered in the input grid. A real-world experiment can be handled in this way: first, filter out the point cloud of the object in the global coordinate system, then calculate the center and the bounding box. Input the grid into the network consisting of each occupancy value of the points within the bounding box in XYZ order (as the network takes 323232*1 occupancy values as input). Note that the table points inside the bounding box are needed to be included.