Siamese Networks or SIFT Approach?

I'm developing a image similarity detection system(with few-shot learning approach) in Python for our project and I'm indecisive about which one should I use: Siamese or SIFT. Siamese is directly used for image similarity(as it's similarity function consisting) and where I search about the few-shot learning aproach, I encounter the Siamese, however SIFT seem a high-level and gold standart approach to apply similarity concept on images, as its affine distortion, scaling, rotating etc. invariancy properties. Which one should I use?