07. R-CNN

R-CNN Outputs

The R-CNN is the least sophisticated region-based architecture, but it is the basis for understanding how multiple object recognition algorithms work! It outputs a class score and bounding box coordinates for every input RoI.

An R-CNN feeds an image into a CNN with regions of interest (RoI’s) already identified. Since these RoI’s are of varying sizes, they often need to be warped to be a standard size, since CNN’s typically expect a consistent, square image size as input. After RoI's are warped, the R-CNN architecture, processes these regions one by one and, for each image, produces 1. a class label and 2. a bounding box (that may act as a slight correction to the input region).

1. R-CNN produces bounding box coordinates to reduce localization errors; so a region comes in, but it may not perfectly surround a given object and the output coordinates (x,y,w,h) aim to perfectly localize an object in a given region.
2. R-CNN, unlike other models, does not explicitly produce a confidence score that indicates whether an object is in a region, instead it cleverly produces a set of class scores for which one class is "background". This ends up serving a similar purpose, for example, if the class score for a region is Pbackground = 0.10, it likely contains an object, but if it's Pbackground = 0.90, then the region probably doesn't contain an object.