Conditional Iteration Updated Random Forests for Unconstrained Facial Feature Location

Facial feature location is important for face analysis in computer vision. In order to enhance accuracy and efficiency in unconstrained environment with various head poses, background, occlusion, illumination and make-up, we propose a conditional iteration updated random forest approach for precise facial feature location. In order to eliminate occlusion and background noise, positive/negative facial patches are classified from a facial area, firstly; then, initial facial feature positions are detected by the trained initial conditional probabilistic model under the condition of 25 head pose models in three local sub-regions; the multiple leaves of random forests are updated using the on-line learning in an iterative way based on the detected facial feature initial position and facial error deformation, which include patches’ classes, head pose probabilities, face deformation model and face error offset model; finally, a conditional weighted sparse voting method is introduced into the interactive regression model for voting final precise facial feature locations. Experiments on the challenging AFW, LFW, and Pointing’04 datasets demonstrated that the average accuracy of our approach has reached 95% with only 0.15 mean errors for facial feature location in unconstrained environment.