Figure 12
From: Multi-pose lipreading and audio-visual speech recognition
Lipreading performance associated to frontal and lateral views. This figure shows the loss of performance in visual-only experiments associated to non-frontal poses and the existence of a train/test pose mismatch when no pose normalization is applied to the visual stream. The first column corresponds to the frontal (F) sequences tested on the frontal classifier. The second column to the lateral (L) sequences with the corresponding lateral classifier. The last column to the lateral sequences directly tested on the frontal classifier, without the proposed pose normalization.