How do I reproduce the result of Table 3 where you are running all benchmark on different types of vision backbone?