The R-CNN family

R-CNN, Fast R-CNN, Faster R-CNN, and Mask R-CNN

Traditional Approach

• Region Proposal Generator: Selective Search , Edge Boxes

• Image descriptors: histogram of oriented gradients (HOG)

• Classifier: Support Vector Machine

R-CNN

1. Extract features using a CNN network

Fast R-CNN

Faster R-CNN builds a network that has only a single stage:

1. Input image is fed into a pretrained CNN network to get a feature map

2. Input image is also used to propose $n$ region proposals

3. Project the region proposals on the feature map yields $n$ region of interest (ROI)

4. ROI pooling layers are used to extract a fixed-length feature vector from those ROIs

Fast R-CNN is faster than R-CNN because:

1. Faster R-CNN shares computations (i.e. convolutional layer calculations) across all proposals (i.e. ROIs) rather than doing the calculations for each proposal independently.

2. Fast R-CNN does not cache the extracted features and thus does not need so much disk storage compared to R-CNN, which needs hundreds of gigabytes.

Faster R-CNN

1. Changes the selective search portion in Fast R-CNN into a Regional Proposal Network.

Mask R-CNN

1. Replaces the region of interest pooling layer with the region of interest (RoI) alignment layer

2. More suitable for pixel-level prediction (semantic segmentation)

Resources

14.8. Region-based CNNs (R-CNNs) — Dive into Deep Learning 1.0.0-beta0 documentation

Faster R-CNN Explained for Object Detection Tasks | Paperspace BlogPaperspace Blog