Regression-based schemes have proven effective for locating landmarks on images. Most previous approaches either predict the positions of all points simultaneously, or use regressors that predict individual points combined with a global shape constraint. The former can be efficient, but such models tend to be less robust. Conversely, Random Forest (RF) voting methods for individual points have been shown to be robust and accurate, but can lead to very large models. We explore the continuum between these two approaches by training RF regressors to predict subsets of points.
Multi-point regression voting was implemented within the Random Forest Regression Voting Constrained Local Model frame- work and evaluated on clinical and facial images. Significant model size reductions were achieved with little difference in accuracy. The approach may therefore be useful where high numbers of points, and limitations on memory or disk space, make single-point models impractically large.
