Add reference model that uses semantic map

[gzuidhof]

This adds a reference model implementation that uses the semantic map as additional input (which is not easily possible in the kernel right now).

Class_names =  ['animal', 'bicycle', 'bus', 'car', 'motorcycle', 'other_vehicle', 'pedestrian', 'truck']
Average per class mean average precision =  0.10458624970399849
('animal', 0.0)
('bicycle', 0.014827256648354046)
('bus', 0.07114362365576843)
('car', 0.4644436416741515)
('motorcycle', 0.031743736796870606)
('other_vehicle', 0.240433992246761)
('pedestrian', 0.002291803782829482)
('truck', 0.011805942827252867)