Fast R-CNN是在R-CNN的基础上进行的改进，大致框架是一致的。总体而言，Fast R-CNN相对于R-CNN而言，主要提出了三个改进策略：
1. 提出了RoIPooling，避免了对提取的region proposals进行缩放到224x224，然后经过pre-trained CNN进行检测的步骤，加速了整个网络的learning与inference过程，这个是巨大的改进，并且RoIPooling是可导的，因此使得整个网络可以实现end-to-end learning，这个可以认为是Fast R-CNN相对于R-CNN最大的改进之处。
2. 采用了Multi-task loss进行边框回归，这个在R-CNN中也有这方面的实验。
3. 利用了截断的奇异值分解（Truncated SVD for faster detection）加速了网络。(Fast r-cnn is an improvement based on r-cnn, and the general fr a mework is consistent. In general, fast r-cnn, compared with r-cnn, mainly proposes three improvement strategies:
1. Proposed roipooling to avoid scaling the extracted region proposals to 224x224, and then accelerated the learning and information process of the whole network through pre trained CNN detection steps. This is a huge improvement, and roipooling is derivable, so that the whole network can achieve end-to-end learning, which can be considered as fast r-cnn relative to r-cnn- CNN's biggest improvement.
2. Multi task loss is used for border regression, which is also tested in r-cnn.
3. The truncated SVD for fast detection is used to accelerate the network.)