The matrix g  can have CV_32F or CV_64F type,  but g.at uses only float template.
This fix adds specialization for double type.