A new perspective on face detection utilizing language modeling

Most traditional face detection models rely heavily on manually designed prior knowledge. Moreover, for other related facial visual tasks (such as facial attribute analysis, facial pose estimation, etc.) may require additional models or complex integrations. Based on the generalized features of Pix2Seq for visual tasks, this paper studies its application in face detection, thus providing a universal language interface for facial visual tasks. We define face detection as a language modeling task based on pixel input. In order to speed up the training data, we used DETR model code to implement the Pix2Seq model algorithm for transforming from image to sequence, and add data augmentations. To further compress the model and reduce the computational load, we choose to delete the FFN. By training Pix2Seq on the large-scale WIDER Face dataset, we report the latest results on the WIDER Face validation set.