Useless and noise information occupies large amount of big data, which increases our difficulty to extract worthy information. Therefore outlier detection attracts much attention recently, but if two points are far from other points but are relatively close to each other, they are less likely to be detected as outliers because of their adjacency to each other. In this situation, outliers are hidden by each other. In this paper, we propose a new perspective of hidden outlier. Experimental results show that it is more accurate than existing distance-based definitions of outliers. Accordingly, we exploit a candidate set based hidden outlier detection (HOD) algorithm. HOD algorithm achieves higher accuracy with comparable running time. Further, we develop an index based HOD (iHOD) algorithm to get higher detection speed.