The pipeline of our approach is shown in Figure 1. We first develop a scheme for structured feature learning, aiming to integrate local spatial label patterns into the feature space. Then, we feed the resulting features to train two boosted tree classifiers, as shown in Figure 2. The first boosted tree classifier is used to obtain context distance features. The resulting context distance features, together with structured features, are used to train the second boosted tree classifier which is adopted for testing image segmentation. Figure 3 shows the growing of a single tree as well as an illustration of why structured features are useful.
Figure 1: Pipeline of our approach.
Figure 2: Illustration of the construction of context distance features. Left panel illustrates the pipeline of training two boosted tree classifiers.
Right panel presents the construction of context distance features for a specific pixel, where only top $d_c$ context distance features are preserved.
Figure 3: Illustration of a single tree in the boosted tree classifier. Left panel illustrates why structured features are useful for partitioning patches containing similar filamentary structures into the same leaf node. Right panel shows the function of a single tree, which corresponds to a tree in boosted tree classifier I when input contains only structured features and corresponds to a tree in boosted tree classifier II when input contains both structured and context distance features.
|