This is the expected bounding box (2d tight):
This is the ROS topic subscribed one (same camera on the right), just shown at different frame.
It seems that the ROS topic publishes a much tighter frame that shrinks a lot (it also happens for the other objects). Can I force the ROS topic to somehow publish the other ‘tight’ version, or the ‘loose’ version.
Also on a different note, the instance segmentation also doesn’t correspond - in synthetic view it is based on the assigned labels, in ROS topic it’s for each part (aka. non-parsed). I’m assuming ROS topics cannot be published for the MotionVector and Normals images and they have to be generated anew.