Master Thesis: Structured learning with random forests

Recently there has been a significant progress in human pose estimation from range maps and fully 3d data. Thanks to decision forest, now people can infer correspondences (either sparse or dense) between a surface model and single depth image/visual hulls. However, most of the time, the inferred correspondences do not respect the intrinsic structure of meshes, i.e., edge connectivity.

In the previous project, we investigated how to regularize/correct forest predictions with MRF/CRF as a spost-processing step. In this project, we would like take one step further, investigating how to merge MRF/CRF framework into forests, first in the prediction stage and then eventually in the training stage. That is, we hope that forests consider the mesh connectivity to learn and to predict (as in the right figure). This goes into the direction of structured learning. Analogous to the structure-learning version of SVM (termed structured SVM), we would like to see if one can achieve similar things with regression forests.