Convolutional Hypercolumns
Last updated
Last updated
Refer to this link
Hypercolumn of a pixel is the vector of activations of all CNN units “above” that pixel. They contain missing information from the first and last layers.
Hypercolumns are something you extract from the network, not build.
The last fully connected layer in a CNN may be too coarse too allow for localization precision. The first layers could be spatially precise but lacks semantic information. Hypercolumns help with improving both since it contains information on the first and last layers.
