April 2018
Intermediate to advanced
334 pages
10h 18m
English
By removing the focus on unnecessary regions of the image R-CNN was faster than normal CNN but still R-CNN was practically very slow, since the number of regions R-CNN was focusing on was high enough for the overall computation to be still expensive.
Spatial Pooling Pyramid networks (SPP-net) were the first attempt to fix this issue. In SSP-net, the CNN representation for the entire image is calculated only once and that is further used to calculate the CNN representation for each of the box regions generated by the selective search approach. This is done by pooling on that section of the convolution representation corresponding to the box region. The section of convolution representation corresponding to ...