R-FCN is more similar to R-CNN than SSD. R-FCN was developed in 2016 by a team, mainly from Microsoft Research, consisting of Jifeng Dai, Yi Li, Kaiming He, and Jian Sun in a paper titled R-FCN: Object Detection via Region-Based Fully Convolutional Networks. You can find the link for the paper at https://arxiv.org/abs/1605.06409.
R-FCN is also based on region proposal. The key difference from R-CNN is instead of starting with 2K region proposal network, R-FCN waits until the last layer and then applies selective pooling to extract features for prediction. We will train our custom model using R-FCN, in this chapter, and we will compare the final results with other models. The architecture of R-FCN is described in the following diagram:
In the preceding figure, an image of a car is passed through ResNet-101, which generates a feature map. Note that we learned...