Object classification is a core element of various robot services ranging from environment mapping and object manipulation to human activity understanding. Due to limits in the robot configuration space or occlusions, a deeper understanding is needed on the potential of partial, multi-view based recognition. Towards this goal, we benchmark a number of schemes for hypothesis fusion under different environment assumptions and observation capacities, using a large-scale ground truth dataset and a baseline view-based recognition methodology. The obtained results highlight important aspects that should be taken into account when designing multi-view based recognition pipelines and converge to a hybrid scheme of enhanced performance as well as utility.