EigenCAM

• @banymuhammadEigenCAMVisualExplanations2021

Summary by Me

Another method for computing Saliency Maps without modifying the architecture of the network, EigenCAM was proposed by Bany et al. \cite{banymuhammadEigenCAMVisualExplanations2021}. EigenCAM uses a combination of an Eigen analysis of the class activated output by projecting it on the input, and a PCA of it to remove unnecessary features from the maps. The Eigen-Saliency map is computed across the network and produces sharper outputs based on the distance (using PCA) from the input image. EigenCAM and Eigen Saliency maps were fused by a point wise multiplication operation.