In this paper, we evaluate the first of two parts of a novel approach for the assessment of residual error probabilities in trained convolutional neural networks (CNN). We consider CNNs for camera image classification, as needed in a safety-critical context, for example in autonomous road vehicles or trains, for the purpose of obstacle detection. The objective of the strategy’s first part is to identify so-called classification clusters of CNNs: these are subsets of the input space, whose elements are all mapped to the same classification result. To this end, we apply a new technique based on mathematical analysis.