Summarizing:

2.5d uses depth estimation for each feature extracted from image. The z is different for each feature.

2d use only coordinates x and y of features extracted from image, z can be computed from a calibration process of the camera. The z is the same for some features.

The depth estimation can be obtained from a stereo pair, from a rgbd sensor or other device.

F. Chaumette and E. Malis are two important experts in this topic. You should read their papers.

Here you have a work about visual servoing based on image using depth for each feature or for the set of features. In addition, you have some references about those important authors.

https://www.researchgate.net/publication/259082588_Guidance_of_Robot_Arms_using_Depth_Data_from_RGB-D_Camera?ev=prf_pub

Conference Paper Guidance of Robot Arms using Depth Data from RGB-D Camera