VL_NNPDIST
 CNN pdistance from target.
VL_NNPDIST(X, X0, P) computes the P distance raised of each feature vector in X to the corresponding feature vector in X0:
Y(i,j,1) = (SUM_d (X(i,j,d)  X0(i,j,d))^P)^(1/P)
X0 should have the same size as X; the outoput Y has the same height and width as X, but depth equal to 1. Optionally, X0 can be a 1 x 1 x D x N array, in which case the same target feature vector in X0 is compared to all feature vectors in X. In that case, however, the DZDX0 are of size of X.
Setting the noRoot
option to true
does not take the 1/P power
in the formula, computing instead
Y(i,j,1) = SUM_d (X(i,j,d)  X0(i,j,d))^P
For example, vl_nnpdist(x, x0, 2, 'noRoot', true)
computes the
squared L2 distance.
[DZDX, DZDX0] = VL_NNPDISTP(X, X0, P, DZDY) computes the derivative of the block inputs projected onto DZDY. DZDX, DZDX0 and DZDY have the same dimensions as X and Y, respectively.
VL_NNPDIST(___, 'OPT', VAL, ...) accepts the following options:

NoRoot
[false
]If set to true, compute the Pdistance to the Pth power.

Epsilon
[1e6]When computing derivatives, quantities that are divided in are lower boudned by this value. For example, the L2 distance is not smooth at the origin; this option prevents the derivative from diverging.

Aggregate
[false]Instead of returning one scalar for each spatial location in the inputs, sum all of them into a single scalar.

InstanceWeights
[[]
]Optionally weight individual instances. This parameter can be eigther a scalar or a weight mask, one for each pixel in the input tensor.