Guided Spatial Transformers for Facial Expression Recognition

Spatial Transformer Networks are considered a powerful algorithm to learn the main areas of an image, but still, they could be more efficient by receiving images with embedded expert knowledge. This paper aims improve performance conventional Transformers when applied Facial Expression Recognition. Based on Transformers’ capacity spatial manipulation within networks, we propose different extensions these models where effective attentional regions captured employing facial landmarks or visual saliency maps. specific information is then hardcoded guide transformations that best fit proposed for better recognition results. For this study, use two datasets: AffectNet and FER-2013. AffectNet, achieve 0.35% point absolute improvement relative traditional Transformer, whereas FER-2013, our solution gets increase 1.49% fine-tuned Affectnet pre-trained weights.