It remains a challenging issue to localize blocked RFID objects. Existing solutions rely on non-line-of-sight (NLOS) signals reflected from environments, which are not reliable and cannot be accurately measured. To eliminate such limitations, this paper develops a new approach (called MetaSight) that localizes blocked objects by modulating NLOS signals via metasurfaces. More specifically, a programmable metasurface is designed such that each reflecting element can dynamically change frequencies and phases of reflected signals. With proper setting of frequencies and phases within a certain range of reflecting elements (i.e., the reflecting window) by a reflection beamforming algorithm, a unique NLOS signal path (called metasurface path or MS path) is established between an object and its reader. By shifting the reflecting window in time domain, multiple MS paths and the corresponding channel coefficients can be obtained sequentially. To make the sequential process fast, both the preamble and the payload of a frame are utilized for channel estimation. Based on the estimated channel coefficients of multiple MS paths, the 3D direction of the object viewed from the metasurface is derived through a 3D direction estimation algorithm. By consolidating the 3D directions from two separate metasurfaces, the object's 3D location is finally determined. MetaSight is implemented as a prototype system. Extensive experiments show that MetaSight can localize a blocked object with an average accuracy of 15 cm.