The short answer is: yes, it could. In what you are describing, there's nothing very new or specific conceptually; it sounds like a standard regression task. Now the problem that you're actually facing is: do you have the data?
Algorithms won't be able to learn the distance between eyes if you don't have the data that it takes. It could be supervised labels (1 distance per image which would be your regression target), reconstruction from depth maps, multi-view estimation etc. There's a number of ways you could do that given the appropriate data.
People focus on algorithms a lot, and that's good. But taking a good look at your data is often as important (if not more).
Now a good example would be in the self-driving car literature. You could start with this blog-post and go through the papers they reference: https://towardsdatascience.com/vehicle-detection-and-distance-estimation-7acde48256e1
There also seems to be some litterature about your eyes examples (https://arxiv.org/pdf/1806.10890.pdf, https://www.sciencedirect.com/science/article/pii/S0165027019301578) so skimming through these papers & the datasets they use could guide to towards answering my question: is there data for this task?