The prevalence of imperfections in data, characterized by uncertainty and imprecision, prompts the need for effective modeling techniques. The theory of belief functions offers a mathematical framework to address this challenge. In this paper, we tackle the problem of calculating the mean distance between elements of the same class, especially when class membership is uncertain and imprecise. Leveraging belief functions and a notion of similarity between elements, we propose a solution and validate its efficacy through experimental evaluations. The proposed method proves effective when labels exhibit low imprecision, whereas unsupervised methods may be more effective for labels closer to complete ignorance.