Mask RCNN implementation in PyTorch
Faster RCNN works in two phases:
- Region proposals
- Classifying regions
A region is an area of the original picture which might contain an object. Also known as Region Of Interest (RoI)
These are the most important aspects of an RCNN. They are also a source of bottlenecks.
RCNN used a large number of region proposals by running it through a variety of category independent region proposal algorithms. These regions are then passed through a CNN.
Fast RCNN is an improvment over RCNN. Instead of running the region proposal algorithms over the underlying image, the algorithms are run over a feature map. This feature map is obtained by passing the image through the Convlutional layers of any CNN. Fast RCNN is computationally less expensive when compared to RCNN.