Colorization problem is considered.
A fixed number of features is associated with each vertex (color, area, etc.) and with each edge (length of common border, color contrast, etc.).
For input graph (an image to be colored) the trained model is supposed to specify color for each vertex (color region) of the input graph.
Classic regression algorithms (svm, decision trees, etc.) take fixed sized vectors as input and return fixed sized vectors.
Is there a model that naturally handles varying number of vertices / edges?
Is there an approach to adapt fixed structure models (MRF, factor graphs, etc.) to handle this particular case?