In 2015, Fast R-CNN was proposed to remedy the speed problems of R-CNN. In this method, the main change is where we get proposal regions in the pipeline. Instead of getting them directly from the input image, we first run the entire input image through a CNN and extract the generated feature map close to the end of the network. Next, again using a region-proposal method, candidate regions are extracted from this feature map in a similar manner to R-CNN.

Getting proposals in this way helps reuse and share expensive convolutional computations. The fully connected layers further down in the network that will classify, and additionally localize, only accept fixed-size input. For this reason, the proposed regions from the feature map are warped to a fixed size using a new layer called RoI pooling (discussed further in the next section). RoI Pooling resizes the regions into the size needed by the last FC layers. The whole process is shown in the following diagram:

The comparison of R-CNN and FastRCNN shows that the latter is about 10 times faster at training time and nearly 150 times faster at test time (when using a VGG architecture as the main CNN).