Ask HN: How to recognise image when parts of it may be occluded?

Problem: Given a frame from a mobile camera's video stream, I need to pick a reference image (from a corpus of images) that the given frame most closely resembles. Input: Mobile frame, Output: file name of reference image. Sort of like a mobile version of TinEye or Google image search, except that the corpus is at most a few thousand images.

I can constrain the problem further. I can make sure that the mobile camera is positioned at about the same angle that the reference images were taken; but the mobile used to capture the reference may not be the same as the one 'in the field'; given the differences in resolution, lighting, different types of mobile camera specs etc, the images are bound to differ.

Second, it is possible that some part of the image to be identified is occluded (like a hand or a paper weight).

This needs to work on Android (which I am new to), and I don't have much of a background in image recognition. Ideally, if there is a way to fingerprint/hash the image in a way that is resistant to occlusion, rotation etc, and perhaps even the angle it is viewed from, that would be nice. I have thought of using Unity+Vuforia, OpenCV, perceptual hashing (phash) but I'd be very grateful for comments from experienced devs. Thank you all.