Figure 15
From: Multi-pose lipreading and audio-visual speech recognition
Lipreading performance with pose normalization in the DCT space. This figure shows the performance of a frontal classifier when the input corresponds to a lateral view of the speaker and the visual features have been normalized to a frontal pose in the DCT space. GLR outperforms LLR because the patches corresponding to high frequencies (containing the details of the image) can not be matched by a linear mapping.