We propose a deep learning approach to perform road-detection in LIDAR scans, at the point level. Instead of processing a full LIDAR point-cloud, LIDAR rings can be processed individually. To account for the geometrical diversity among LIDAR rings, an homothety rescaling factor can be predicted during the classification, to realign all the LIDAR rings and facilitate the training. This scale factor is learnt in a semi-supervised fashion. A performant classification can then be achieved with a relatively simple system. Furthermore, evidential mass values can be generated for each point from an observation of the conflict at the output of the network, which enables the classification results to be fused in evidential grids. Experiments are done on real-life LIDAR scans that were labelled from a lane-level centimetric map, to evaluate the classification performances.