Several local region partition methods for Speeded Up Robust Features (SURF) descriptors were researched to reduce their dimensions and increase the matching speeds and robustnesses of SURF based matching algorithms. With reference to existing local region partition methods of Scale Invariant Feature Transform (SIFT) and SURF descriptors, the local regions were divided into grids(original SURF), triangles, and sectors. First, the influences of scale change and rotation of the image on the matching performance of SURF descriptors were analyzed. Then, a method to construct the SURF descriptors with local region partitions in triangles and sectors were proposed, and matching experiments were performed. The SURF descriptors with different local region partitions were compared. The experiment results show that the performance of the sector partition based SURF descriptor is better than those of triangle partition and grid partition (original SURF) based SURF descriptors. The performance of SURF descriptors with 6-sector partition, 8-sector partition, 12-sector partition and triangle partition is better than that of original one, and the dimensions of these new descriptors are 40, 32 , 16 and 32 lower than that of original SURF (64 dimensions).