Updated the lambda coefficient to the one suggested in the original paper (http://www.robots.ox.ac.uk/~joao/publications/henriques_tpami2015.pdf, page 11, Table 2) which works (for me) better then the previous default one.