Mastering Computer Vision with TensorFlow 2.x

Graph-based segmentation

There are many graph-based segmentation methods available, but the one described here for R-CNN is the one introduced by Pedro Felzenszwalb and Daniel Huttenlocher in a paper titled Efficient Graph-Based Image Segmentation. The paper can be read at http://people.cs.uchicago.edu/~pff/papers/seg-ijcv.pdf.

This method involves representing an image as a graph (a detailed explanation is given in the Overview of GNN section in this chapter) and then selecting edges from the graph where each pixel is linked to a node in the graph and connected to neighboring pixels through edges. The weights on an edge represent dissimilarity between pixels. The segmentation criteria are based on the degree of variability of neighboring ...