Indoor Scene Semantic Labeling Using Random Decision Forests

Recently scene labeling appeals to many researchers with the developing of theintelligent sensors and machine learning algorithms. In the field of image understanding,camera is used to capture scene frame with the purpose of generating a higher leveldescription to raw data, and then the decision system can be more efficient.In this paper we focus on indoor scene semantic labeling towards each pixel. Our workshows how to make use of random decision forests and get the accurate semantic labels,which will benefits to service robotic systems, intelligent housing systems, emergencyresponse systems etc. In different stages of labeling, we give different solution based onrandom decision forests:1. Information fusion with color image and depth image. Our model firstly acquiresRGB color frame and depth frame simultaneously using RGB-D camera, and learns the textonmodes within small local patches. Color textons and depth textons are fused in the nodes ofrandom decision forests, and then provide primary and local scene labeling.2. Semantic consistency of multi-classes. We adopt random decision forest model as asemantic consistency optimized model to handle indoor scene multiple classificationproblems. This optimization is based on the fact that relative position of two classes candescribe the labeling rationality in some way. In this paper we discuss two differentextensions to position and classes’relations to improve the accuracy of labeling.3. Global optimization of scene labeling. We use a random decision forest model inedge detection, and integrate it in Markov Random Field model, in which the contourinformation is regards as pairwise potential energy. Experimental results show this approachescan effectively suppress noise to generate more reliable results.