In Lowe's paper Section 4.1 the ratio of principal curvatures using the Hessian Matrix is used to eliminate points that may belong to an edge.
The paper does not specify whether the Hessian matrix is computed on the original image or the DoG. In Rob Hess's implementation it is applied on the DoG at the specified octave & interval.
My question is why is the Hessian Matrix applied on the DoG to identify points that are potential edges?
I understand from Utkarsh's tutorial that the Hessian is supposed to be similar to Harris corner detection. The difference is that the Hessian matrix is a second-order derivative. So I concluded that the Hessian should be applied to the original image.
Can you point me to relevant resources that explain how the Hessian is used?