Sylvain: without a lot of data, the measured classification rates are not well estimated (high variance). By multiplying the measurement points and calculating the average of the accuracy maps, a better estimate of the average is obtained. So better performance measurement with leave 2 out and more statistical power. But obviously it is a compromise between a well trained classifier and a well estimated performance measure.
Sylvain: without a lot of data, the measured classification rates are not well estimated (high variance). By multiplying the measurement points and calculating the average of the accuracy maps, a better estimate of the average is obtained. So better performance measurement with leave 2 out and more statistical power. But obviously it is a compromise between a well trained classifier and a well estimated performance measure.