This document summarizes and compares R-CNN and Fast R-CNN object detection models. Fast R-CNN improves on R-CNN by having a single-stage training pipeline rather than multi-stage, and by feeding the whole image through the convolutional network rather than individual regions of interest, improving speed. It also uses ROI pooling rather than region warping and adds a multi-task loss function. The document discusses techniques used in Fast R-CNN like image pyramids, scale invariance, and truncated SVD to reduce model parameters.