calculate weights inverse to the distances from the specified observation.
for calculation of a shape model by averaging the observations neighbouring the configuration in question, it is necessary to calculate weights by similarity.
proc.weight(data, number, ref, report = TRUE, reg = 0, log = FALSE, mahalanobis = FALSE, weightfun = NULL)
data |
array containing landmark configurations |
number |
integer: how many of the neighbours are to be involved. |
ref |
integer: position in the array that is used as reference. |
report |
logical: require report about name of the reference. |
reg |
numeric: regularise mahalanobis distance by adding reg to the diagonal of eigenvalues of the covariance matrix. |
log |
logical: use the logarithm of the distances. |
mahalanobis |
logical: use mahalanobis distance. |
weightfun |
custom function that operates on a vector of distances (see examples) and generates weights accordingly. |
distances of zero will get a weight of 1e12 (this is scaled to all weights summing to one), thus weights for observations further away are converging to zero.
data |
dataframe containing id, procrustes/mahalanobis distance and weight according to the reference |
reference |
returns observations' names if available |
rho.all |
dataframe containing distances to references of all observations |
if (require(shapes)) {
proc <- procSym(gorf.dat)
##get weights for the the four specimen closest to the first observation.
weights <- proc.weight(proc$rotated,4,1)
##estimate the first specimen by weighted neighbour shapes.
estim <- proc$mshape*0;
for (i in 1:4)
{estim <-estim+proc$rotated[,,weights$data$nr[i]]*weights$data$weight[i]}
### visualise
plot(estim,asp=1)## show estimation
points(proc$rotated[,,1],col=3)##show original
## use a gaussian smoother to compute weights using a bandwidth of 0.05
gaussWeight <- function(r,sigma=0.05) {
sigma <- 2*sigma^2
return(exp(-r^2/ sigma))
}
weights <- proc.weight(proc$rotated,4,1,weightfun=gaussWeight)
}Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.