Common visual pattern discovery via spatially coherent correspondences