Applying deep residual architectures like ResNet-50 and ResNet-101 to the Faster R-CNN framework requires addressing the absence of standard hidden fully connected layers found in models such as VGG-16. This is resolved by adopting the Networks on Conv feature maps (NoC) approach. Convolutional layers with a stride on the input image no greater than 16 pixels (specifically conv1, conv2_x, conv3_x, and conv4_x, totaling 91 convolutional layers in ResNet-101) compute shared full-image feature maps with a total stride of 16 pixels.

The shared feature representation is used by both a Region Proposal Network (RPN), which generates 300 proposals, and a Fast R-CNN detection network. Region of Interest (RoI) pooling is inserted directly prior to conv5_1. The remaining convolutional layers in conv5_x are applied independently to each RoI feature, serving the functional role traditionally filled by fully connected layers, before terminating in sibling classification and box regression branches.

To control memory consumption during Faster R-CNN fine-tuning, Batch Normalization (BN) layer statistics (means and variances) are precomputed on the ImageNet training set and kept fixed. Consequently, the BN layers act as linear activations with fixed offsets and scales rather than tracking dynamic mini-batch statistics during detection training.