Several common estimation problems where one is interested in estimating models from paired input/output observations can be considered as being non-statistical problems: learning preferences from a single user rather than a population, inverse problems where the physical model can be trusted, .... In this case, one can easily question whether imperfection in observations and the ensuing estimation should be treated as a statistical problem. In particular, one could challenge the need to model imperfect observation as probabilistic noise, and the fact of considered expected errors as a measure of estimation quality. With these ideas in mind, this work discusses an alternative view, where imperfection is modelled by uncertainty theories that accommodate imprecision by extending set-valued representation, and where estimation is mainly performed by using information fusion tools building upon standard union and intersections rather than averaging and counting considerations.