CSE 573 Final Project: Semantic Labeling

Automatic image segmentation and semantic labeling is one of the most important and challenging problems in computer vision. Labeling could be done at various quantization levels – pixel level, super-pixel level, or group of superpixel (segment) level. Each approach has its advantages and disadvantages. In this paper, we explore the superpixel level semantic labeling of images using supervised machine learning techniques. The labeling was performed using several low-level features – location, color, and texture of super-pixels. Four different classifiers were trained and their performances were compared. These classifiers are – support vector machines, logistic regression, adaptive boosting, and conditional random fields. Evaluation of these classifiers was done using large dataset of images with 8 kinds of labels. With a small set of primitive features, a very encouraging performance (70% accuracy) was achieved.