Facial Feature Reconstruction using Structure from Motion

Abstract

Structure from Motion suggests that an object or scene’s three-dimensional structure can be determined from its observed two-dimensional motion. Human efforts, manifested in computer algorithms, try to mimic the enormous power of the visual processing capabilities of the human brain. We present an algorithm to estimate structure, using the Unscented Kalman Filter, from the motion of point-wise features, produced by the Kanade-Lucas-Tomasi feature tracker. The algorithm is evaluated critically against an extensive set of motion sequences, with special attention paid to facial feature reconstruction.