3D detection depth is inconsistent

[gadorneles]

We use the center point of the bounding box to get the depth information on the object, which can be very inconsistent. I think the best way of solving this is using a mask in img2world and getting the median of all points.

[gadorneles]

@LuisMilczarek talked about forming a cluster in the point cloud of the object, which is also a path we can take. Although it's most likely more demanding than just using masks.