Multiple-Layer Classifier with Label Correction for Semantic Segmentation
Lavinia Eugenia Ferariu, Simona Caraiman · 2018
Semantic segmentation (SS) provides the meaning of visual scenes, thus being a key stage for navigation and environment's perception. This paper presents a solution for SS compatible with assistive wearable systems equipped with color and depth cameras. In order to ensure a compact and robust description of input color-based images, both 2D and 3D features are extracted at superpixel level, after correcting the displacements of the camera by means of adequate rectifications. Random Forests (RF) are called for solving the classification problem. In this context, this paper introduces a multilayer RFbased classifier, including a separate layer for label correction. Additional two other correcting methods are proposed for the first layers of the classifier, i.e. a fast method investigating the majority label around each object, and several customizations of the graph cut algorithm using convenient cue weights. The performance of the suggested approach is experimentally verified on diverse urban street scenes.