Geodesic pixel neighborhoods for multi-class image segmentation

The problem of multi-class image segmentation is traditionally addressed either with a Conditional Random Field model or by directly classifying each pixel. In this paper we introduce a new classification framework built around the concept of pixel neighborhoods. A standard unary classifier is used to recognize each pixel based on features computed from the image and then a second classifier is trained on the predictions from the first one summarized over the neighborhood of each pixel by a new histogram feature. We define a local and a global pixel neighborhoods which adapt to the image structure by making use of the geodesic distance defined over image intensities. We evaluate our model on three challenging datasets and show that our model is able to capture both local and global context relations. We compare our method to two strongly related, well known methods and show increased performance.