There are successful solutions available based on eCognition, for instance:
http://www.trimble.com/imaging/pdf/TM18_FINAL_INPHO.pdf

eCognition offers the possibility to translate your mind model (why you see buildings in our images) into computer understandable code (= rule set). The eCognition Developer is the environment to create rule set. A good rule set strategy for your issue is

(1) create small objects only on the spectral information with a 'multiresolution segmetation' 

(2) merge the objects with an ‘image object fusion’ to fuse all objects with similar color AND a defined common border AND similar heights. For this you can use the free customized algorithm (= rule set) from here: http://community.ecognition.com/home/Customized%20Segmentation.zip/view

(3) classify all elevated objects 

(4) calculate the best threshold to separate the elevated objects into Vegetation and Building. For this you can use a free customized algorithm approach (=rule set) from the eCognition Community (called automated vegetation mapping, you have to use the seach function of the Community , I can post only 2 links)

(5) reshape the building object into rectangular objects with the free customized algorithm 'building generalization' from the Community (I can post only 2 links) 

Please note, you can combine existing rule sets in your rule set.

In the eCognition Community you can find Guided Tours and you will get tips and tricks. 

For beginners, I recommend to use the free eCognition Trial version. eCognition Developer Trial version allows  non time limited full software access with the following restrictions: Disabled, in this demo: export functions, saving of rule sets, workspace environment. 

There are different price level for the eCognition software, depends on the profic focus of your company (COM, NPO and EDU). But you can also check Planet Action: Planet Action donated more than 1,000 satellite images alone in 2011, and program partner Trimble donated its eCognition image analysis software licenses to aid users in their image processing tasks—to date, more than 100 eCognition Developer licenses have been donated to more than 50 projects in 43 countries.