If you want to build it yourself from beginning to end, as a beginner, I suggest you don't start from these. In order to better understand machine learning, you should start with simple perceptron, SVM, RBF and Bayesian network, and then learn some knowledge in CNN and computer vision (receptive field, image convolution, etc.). ), and then combine the initial selective search with the later regional proposal method, and then start from RCNN-Fast RCNN-Fast RCNN.

Of course, if we only imitate and reproduce, and use the structure given in this paper and software with neural network library (pytorch or TensorFlow) to realize it, the difficulty will be greatly reduced, but it will still be more difficult. After all, what the paper gives is not comprehensive, and various problems will still be encountered. I suggest that it is best to find the relevant project code from github, and first set up the environment of your machine (GPU, various necessary libraries, etc.). ), and then use other people's code to run (at least demo), learn bit by bit, fully understand before doing your own work (you can transfer directly if you don't want to build).