Fast R-CNN has a lot of similarities with R-CNN models. They all use selective search algorithms to select regions. There are only two different points between R-CNN and fast R-CNN.

Since the CNN models only accept a certain size input, R-CNN models choose to crop and warp the different regions so that they can be calculated. However, after this process, the picture might lose information or become another shape. In order to solve this problem, people add a ROI pooling layer in the fast R-CNN model.

Since ROI pooling can accept any size of input, warp and crop can be avoided to increase the accuracy.

Another different point is that fast R-CNN chose SoftMax instead of SVM to classify.