I used to obtain reasonable estimates like this: get everyone involved together and make a big list of things that need to be done. Everyone independently writes down how hard each thing seems (usually on a relative scale like 1 thru 5). Put the estimates in a spreadsheet and project it onto a wall. For any item where there's significant variance, argue about it until the variance dies down, or break the thing up into smaller tasks and re-estimate. Then take the mean of the estimates for each item. We'd insist on everyone playing along, even if they didn't know very much. The more data points the better.
Nowadays, though, I use the much better algorithm "it's done when it's done". It goes well with the software process, "program our asses off".