Channel frequency response (CFR) is a fine-grained location-specific information in WiFi systems that can be utilized in indoor positioning systems (IPSs). However, CFR-based IPSs can hardly achieve an accuracy in the centimeter level due to the limited bandwidth in WiFi systems. To achieve such accuracy using WiFi devices, we propose an IPS that fully harnesses the spatial diversity in Multiple-Input-Multiple-Output (MIMO) WiFi systems, which leads to a much larger effective bandwidth. The proposed IPS obtains CFRs associated with locations-of-interest on multiple antenna links during the training phase. Then, in the positioning phase, the IPS captures instantaneous CFRs from an unknown location and computes the resonating strength between the instantaneous CFRs and those obtained in the training phase. The IPS compensates the residual synchronization errors in the calculation of resonating strengths. Extensive experimental results in a typical office environment demonstrate that the proposed IPS achieves true positive rates of 99.93% and 100%, and only incurs false positive rates of 1.56% and 1.80% under the line-of-sight (LOS) and non-line-of-sight (NLOS) scenarios, respectively, with a single pair of WiFi devices. Moreover, results with 0.5cm measurement resolution show an accuracy of 1-2cm achieved by the proposed IPS in the NLOS scenario. To the best of our knowledge, this is the first work that can achieve 1-2cm accuracy for NLOS indoor localization using WiFi.