Classic crowd-replication techniques:80 to 100 extras were photographed in different separate passes. Then they were duplicated in Post Production, An immense amount of Roto/Tracking were done to match separate crowds to the final composite. In some cases a detailed color correction is done on each group of crowd to get a better output. Another classic method is to shot different blocks of crowd in different position with blue screen behind or along side them and composite them to clean plate but this method got lot of limitations like the camera must be locked off and the whole composition must be small.
Example of this can be found here: http://www.bftr.com/Pages/movies/vendetta_prisonhi.mov
Shadow-matting and color correction are the common challenges faced on crowd replication when done through classic techniques.
In simple words: shoot smaller crowds and composite them altogether with their digital extras (their own duplicates).
But these days...Industry is moving towards more to AI based crowds which can think and interact with environment around them; then only solution for them is think 3D! In beginning they were using Particles for generating crowd but this technique found difficult due computational difficulties so they began to develop huge scripts which drives the animation which later gave birth software called Massive which is very much known for its crowd replication in LORT! There are other proprietary tools and plug-ins for creating crowd replication today.
Check these links for more information:
And finally check this sexy crowd ;)
Man this is hell lot bikini girls, girls, girls I mean lot of them ;-)