Semantic perception can provide autonomous robots operating under uncertainty with more efficient representation of their environment and better ability for correct loop closures than only geometric features. However, accurate inference of semantics requires measurement models that correctly capture properties of semantic detections such as viewpoint dependence, spatial correlations, and intra- and inter-class variations. Such models should also gracefully handle open-set conditions which may be encountered, keeping track of the resultant model uncertainty. We propose a method for robust visual classification of an object of interest observed from multiple views in the presence of significant localization uncertainty and classifier noise, and possible dataset shift. We use a viewpoint dependent measurement model to capture viewpoint dependence and spatial correlations in classifier scores, showing how to use it in the presence of localization uncertainty. Assuming a Bayesian classifier providing a measure of uncertainty, we show how its outputs can be fused in the context of the above model, allowing robust classification under model uncertainty when novel scenes are encountered. We present statistical evaluation of our method both in synthetic simulation, and in a 3D environment where rendered images are fed into a Deep Neural Network classifier. We compare to baseline methods in scenarios of varying difficulty showing improved robustness of our method to localization uncertainty and dataset shift. Finally, we validate our contribution w.r.t. localization uncertainty on a dataset of real-world images.