We introduce a novel statistical approach that quantifies, for the first time, the amount of colocalization of two fluorescent-labeled proteins in an image automatically, removing the bias of visual interpretation. This is done by estimating simultaneously the maximum threshold of intensity for each color below which pixels do not show any statistical correlation. The sensitivity of the method was illustrated on simulated data by statistically confirming the existence of true colocalization in images with as little as 3% colocalization. This method was then tested on a large three-dimensional set of fixed cells cotransfected with CFP/YFP pairs of proteins that either co-compartmentalized, interacted, or were just randomly localized in the nucleolus. In this test, the algorithm successfully distinguished random color overlap from colocalization due to either co-compartmentalization or interaction, and results were verified by fluorescence resonance energy transfer. The accuracy and consistency of our algorithm was further illustrated by measuring, for the first time in live cells, the dissociation rate (k_d) of the HIV-1 Rev/CRM1 export complex induced by the cytotoxin leptomycin B. Rev/CRM1 colocalization in nucleoli dropped exponentially after addition of leptomycin B at a rate of 1.25×10⁻³ s⁻¹. More generally, this algorithm can be used to answer a variety of biological questions involving protein-protein interactions or co-compartmentalization and can be generalized to colocalization of more than two colors.