Latent Hierarchical Part Based Models for Road Scene Understanding

Road scenes can be naturally interpreted in terms of a hierarchical structure consisting of parts and sub-parts, which captures different degrees of abstraction at different levels of the hierarchy. We introduce Latent Hierarchical Part based Models (LHPMs), which provide a promising framework for interpreting an image using a tree structure, in the case when the root filter for non-leaf nodes may not be available. While HPMs have been developed in the context of object detection and pose estimation, their application to scene understanding is restricted, due to the requirement of having root filters for non-leaf nodes. In this work, we propose a generalization of HPMs that dispenses with the need for having root filters for non-leaf nodes, by treating them as latent variables within a Dynamic Programming based optimization scheme. We experimentally demonstrate the importance of LHPMs for road scene understanding on Continental and KITTI datasets respectively. We find that the hierarchical interpretation leads to intuitive scene descriptions, that is central for autonomous driving.