For a tree to stand high, its roots must go deep. For any model to perform well, its roots, the features should be carefully chosen and should be analysed in depth. When these features grow in number, there is a requirement to fuse these numerous vectors to a condensed yet informative form. For this purpose, there are many approaches for the feature fusion. To categorize them, we can do on basis whether the fusion technique uses a model or not. Sub-categories may be formed within each category of fusion. Also to evaluate them, it is important to know how we should compare them and by using what means. By summarizing these concepts one may become clear on the data they have, the fusion they have to apply and the metrics they have to use. Our paper tries to address this for any kind of problem. We have considered mainly the input to be of different combinations (multi-modal) of text, audio(or signals), and images(or videos, which are a series of images).