Genotyping accounts for a substantial part of the cost of genomic selection (GS). Using both dense and sparse SNP chips, together with imputation of missing genotypes, can reduce these costs. The aim of this study was to identify the set of candidates that are most important for dense genotyping, when they are used to impute the genotypes of sparsely genotyped animals. In a real pig pedigree, the 2500 most recently born pigs of the last generation, i.e. the target animals, were used for sparse genotyping. Their missing genotypes were imputed using either Beagle or LDMIP from T densely genotyped candidates chosen from the whole pedigree. A new optimization method was derived to identify the best animals for dense genotyping, which minimized the conditional genetic variance of the target animals, using either the pedigree-based relationship matrix (MCA), or a genotypic relationship matrix based on sparse marker genotypes (MCG). These, and five other methods for selecting the T animals were compared, using T = 100 or 200 animals, SNP genotypes were obtained assuming Ne =100 or 200, and MAF thresholds set to D = 0.01, 0.05 or 0.10. The performances of the methods were compared using the following criteria: call rate of true genotypes, accuracy of genotype prediction, and accuracy of genomic evaluations using the imputed genotypes.For all criteria, MCA and MCG performed better than other selection methods, significantly so for all methods other than selection of sires with the largest numbers of offspring. Methods that choose animals that have the closest average relationship or contribution to the target population gave the lowest accuracy of imputation, in some cases worse than random selection, and should be avoided in practice.Minimization of the conditional variance of the genotypes in target animals provided an effective optimization procedure for prioritizing animals for genotyping or sequencing.