Fig. 2

Detailed Architecture of the Faster-RCNN model. The basic feature extraction network Resnet-50 is split into two parts in our model: 1) layers conv1 to conv4_x is used for extraction of shared features (in the shared layers), 2) layer conv5_x and upper layers further extracts features of proposals for the final classification and regression (in the classifier). And the RPN implemented with three convolutional layers generates proposals from the shared feature map