PIXEL BASED OBJECT RECOGNITION USING TEXTONBOOST

Rick van Rooij

Under supervision of dr. Robby T. Tan

Abstract

The goal of the experimentation project discussed in this report is to automatically partition an input image into semantically meaningful regions each labelled with a specific object class. The general framework is based on the work by Shotton et al. which uses low-level information such as colour, location, edge information and a new feature texture-layout filter to label each pixel with the most likely class. Using conditional random field this information is used to generate regions of a single object class resulting in a pixel-wise semantic segmentation of the input image..

System Pipeline

References

• J. Shotton, J. Winn, C. Rother, A. Criminisi (July 2007). TextonBoost for Image Understanding: Multi-Class Object Recognition and Segmentation by Jointly Modeling Texture, Layout, and Context. International Journal of Computer Vision.