EOPSA-FACE: An Encoder-Only DETR for Classroom Face Detection
Pei Wen, Cheng Sun, Shiwen Zhang, Yuansheng Luo, Haowei Huang, Jin Zhang · 2022
Face detection in classroom scenes has always been a challenging problem because of the diversity and difficulty in extracting face features caused by multi-face clustering and multi-scale face changes. The recently proposed end-to-end object detector called DETR uses tr899978ansformer architecture instead of hand-designed components to obtain global interactive attention information of images, which not only simplifies the model structure but also greatly improves the feature interaction ability. Inspired by these works, we proposed an encoder-only DETR for classroom face detection without reducing the accuracy of detection, dubbed EOPSA-FACE, considering the huge amount of computation caused by the encoding and decoding structure of transformer included in the DETR. Firstly, aiming at the deficiency of multi-scale feature fusion, an efficient pyramid squeeze attention block is used to improve the backbone Resnet, so that the model can learn richer multi-scale feature representation. Secondly, considering the decrease of detection accuracy caused by insufficient allocation of positive and negative samples in the DETR, a pseudo intersection over union (Pseudo-IOU) design was introduced to achieve more accurate sample allocation. Extensive experiments on self-built face datasets in real classroom scenarios demonstrate the superiority of EOPSA-FACE.