CRF-based semantic labeling in miniaturized road scenes

This paper presents an approach for the automatic pixelwise labeling of road scenes using a Probabilistic Graphical Model (PGM). The learning stage is based upon Conditional Random Fields (CRFs) and the inference of the semantic classes is relies on Tree-Reweighted Belief Propagation (TRW). The employment of miniaturized images based on superpixels is proposed and validated to achieve real time classification, which is of interest for the integration of scene understanding capabilities into ADAS and autonomous vehicles. The evaluation is carried out using the KITTI ROAD dataset achieving top results in roads with multiple marked lanes and it is publicly ranked among the state of the art.