Reinforcement Learning with TensorFlow

Fast R-CNN

Firstly, it was Fast R-CNN (proposed by Ross Girshick of Microsoft Research in 2015) that suggested the idea of sharing the convolution outputs among different regions of the image:

Fast R-CNN(https://arxiv.org/pdf/1504.08083.pdf) by Ross Girshick

In Fast R-CNN, an input image and multiple regions of interest are given as an input to a CNNs. Pooling of RoI is done to obtain a fixed-size feature map and then sent through fully connected layers (FCs) to obtain a feature vector. The R-CNN has two output vectors per regions of interest which are as follows:

• Softmax probabilities
• Per-class bounding-box regression offsets

Fast R-CNN ...