Using a depth sensing camera like Kinect, I would like to retrieve the position of an predetermined object (e.g. a cup, fork etc so that I would ultimately be able to grab the object). What would be a way to achieve this?
If you want to recognise a predetermined object (so one that you have seen before) in a 3d scene. It seems this would work similar to the 2d case: using features. Only in this case using 3d feature extractors/descriptors. You provide your model (so the object that you have seen before), and find salient features in the model. In your new scene you also perform feature extraction and then compare the features with your model. If they match well enough, you have found your object. Have a look at the pcl tutorials on the subject.
You probably want to do this, check it out. They also have code http://personalrobotics.ri.cmu.edu/projects/moped.php