Indoor localization is important for many location- based services. The fundamental challenge is the high deployment cost for device, infrastructure, and calibration. This paper develops a blind calibration approach for received signal strength (RSS)-based localization. The essential idea is to employ a device that visits each region exactly once in an indoor area to complete a blind data collection process without recording the route, locations, and timestamps. Thus, the key challenge is to cluster the blind training data into groups and extract the key features to identify the location regions. Classical clustering algorithms fail to work as the data naturally appears as non-clustered due to mutipaths and noise. In this paper, an integrated segmentation and subspace clustering method is developed to exploit both the sequential data structure from the blind data collection process and the signal subspace structure due to the segmented propagation environments. Based on real measurements from an office space, the proposed scheme reduces the region localization error by roughly 50% from a weighted centroid localization (WCL) baseline. In addition, the performance is also comparable to k-nearest neighbor (KNN) and support vector machine (SVM) that require labeled data for calibration.