Semantic Scene Segmentation using Random Multinomial Logit

We introduce Random Multinomial Logit (RML), a general multi-class classifier based on an ensemble of multinomial logistic regression models, and apply it to the task of semantic image segmentation. The algorithm is simple, can be trained efficiently, and has near realtime runtime performance. RML combines the desirable properties of multinomial logistic regression, being stable and theoretically sound, with those of bagging, which is noise-resistant and applicable to large feature spaces. As a second major contribution, we describe a feature selection algorithm for RML based on statistical properties of logistic regression that ensures that computation is not wasted on statistically insignificant features. We evaluate RML on an extremely challenging real-world video dataset of traffic scenes with large illumination, perspective, and intra-class variation, as well as on the 20-class VOC 2008 dataset. Comparisons with recent techniques on the latter dataset demonstrate RML as being state of the art, and advancing it in many cases.