Often clean and fast go hand in hand:

use strict;
use warnings;
use Benchmark qw(cmpthese);

srand (1);
my @testPaths = map {
    join '\\', ${['c:', 'd:', 'e:']}[rand 2],
        map {int rand 100}
        0 .. rand 6
} 1 .. 100;

cmpthese (
    -3,
    {
        JDP_c => sub {paths2treeJDP_c (@testPaths)},
        JDP => sub {paths2treeJDP (@testPaths)},
        GF  => sub {paths2TreeGF  (@testPaths)},
    }
);


sub paths2TreeGF {
    my %pTree;

    for my $path (@_) {
        my @parts = split /\\/, $path;
        my $scan = \%pTree;

        $scan = $scan->{shift @parts} ||= {} while @parts;
    }

    return \%pTree;
}


sub paths2treeJDP {
    my $hr = {};
    @{$hr}{@_} = map {{}} @_;
    my $n_repls;
    do {
        $n_repls = 0;
        for (sort {length ($b) <=> length ($a)} keys %$hr) {
            if (/(.*)\\(.*)/) {
                $hr->{$1}{$2} = delete $hr->{$_};
                $n_repls++;
            }
        }
    } while ($n_repls);
    $hr
}


sub paths2treeJDP_c {
    my $hr = {map {$_ => {}} @_};
    my $n_repls;
    do {
        $n_repls = 0;
        for (keys %$hr) {
            if (/(.*)\\(.*)/) {
                $hr->{$1}{$2} = delete $hr->{$_};
                $n_repls++;
            }
        }
    } while ($n_repls);
    $hr
}
[download]

Prints:

        Rate   JDP JDP_c    GF
JDP    625/s    --  -19%  -47%
JDP_c  773/s   24%    --  -34%
GF    1175/s   88%   52%    --
[download]

The _c variant omits the sort for a modest improvement but the more interesting improvement is omitting the need for the delete.