Unknown Unknowns (UUs) are a kind of test data points on which predictive model confidence is high but the prediction incorrect. It is crucial to identify these instances for better understanding the limitation of predictive models and to avoid critical errors. However, existing methods utilize a fixed model to identify UUs in test data, resulting in limited performance and high cost. To address these limitations, we propose a regional candidate selection algorithm that seamlessly combines a deep neural network and humans to identify UUs in visual data. Specifically, we identify several candidate regions in the data space where UUs have high probability of being present. This is achieved by comparing labels learned by a deep network and predictions obtained with the original classification model. Moreover, inspired by active learning, diversity and training loss are utilized to obtain suitable query sequences. We evaluate our method using a publicly available image dataset. Experimental results conducted on this dataset demonstrate the improved performance of the proposed method over different baselines under different conditions and its robustness to noisy labels.