Fast R-CNN

Improvements w.r.t. R-CNNare:

1. ROI Pooling: to extract an equal-length vector from all the proposed regions

2. Single stage model

3. Shares the CNN layers across all regions, thanks to the ROI Pooling.

4. No disk caching

5. Higher accuracy

The feature map from the last convolutional layer is fed to an ROI Pooling layer to extract a fixed-length vector from each region. ROI Pooling splits the region in a grid and applies max pooling on each cell.

For further info, check the ROI Pooling section.

The ROI Pooling output is fed into a FC layer that splits into two branches:

1. Softmax layer to predict class scores

2. FC layer with a regression head to predict bounding boxes

Drawbacks

• Depends on Selective Search -> very slow