The other website posted on HN yesterday was much more accurate in that it had very large error circle and I was within the circle. This one tries to be way to specific which you simply can't do with this technique.
It could be "Improved Improved" if they had a central server constantly polling the nodes in exactly the same fashion as this script does. If you know the distance from your central server to the nodes, you can use the rolling response times you're polling for to establish a baseline for the client to help factor out server load.
Making a fake image request to these servers introduces a lot of very random sources of latency over what a typical ping would involve.