3D-aware Facial Landmark Detection via Multi-view Consistent Training on Synthetic Data

Figure 1 . We plot the landmark annotations labeled by different annotators with different colors in view #1 of (a). Accurate annotation of non-frontal faces with large angles like view #1 is challenging. This is a major problem since small differences between annotated landmarks in view #1, becomes substantially magnified when projected to view #2. Training a system on such datasets could lead to poor landmark detection accuracy, as shown in (b). We address this issue by proposing a 3D-aware optimization module that enforces multi-view consistency. We show the landmark detection improvement in (c). Magnified insets in (b) and (c) are shown in (d). After refined by the proposed 3D-aware learning, the detected facial landmark is better aligned with the identity.