April 2018
Intermediate to advanced
334 pages
10h 18m
English
In earlier models of object classification, CNNs were very slow and computationally expensive. Moreover, being a classification problem success totally depended on the amount of accuracy. Running convolutions in a CNN is done by sliding the window across all the regions at each layer. Thus, more bounding boxes (total number of different regions analyzed by sliding the window) means higher cost of computation.
Region-based convolution neural networks (R-CNN) were the first take on applying selective search approaches to reduce the number of bounding boxes being fed to the classifier. Moreover, selective searches use features of texture, intensity, color, and so on to create possible box locations of ...