Facial Landmark Localization

Localizing facial landmarks by exemplar-based graph matching (EGM).

Despite the fact that the chin area is partially occluded, EGM still accurately locates the facial landmarks. EGM first finds similar exemplars through a RANSAC step. These exemplars are then used to generate (a) candidate positions for landmarks and to learn (b) an affine-invariant shape constraint, where the position of each landmark (e.g., the chin) is modeled as a weighted linear combination of the other landmarks. By combining these two sources, EGM solves a graph matching problem to obtain (c) the optimal landmark positions.



Localizing facial landmarks (a.k.a., face alignment) is a fundamental step in facial image analysis. However, the problem is still challenging due to the large variability in pose and appearance, and the existence of occlusions in real-world face images.

We present exemplar-based graph matching (EGM)[1] framework for facial landmark localization. Compared to conventional algorithms, EGM has three advantages:


This video compares EGM with the exemplar approach[2] on the Helen dataset.

This dataset contains 348 images annotated with the same 29 landmarks as LFPW dataset, from which we trained the landmark detectors and selected similar exemplars.

Download the [Video 26MB].


The Helen dataset and its label used in the paper available at here.