SSFD: A Face Detector using A Single-scale Feature Map

摘要

In this paper, we present a simple but effective face detector (dubbed SSFD), which can localize multi-scale faces. Unlike other multi-scale feature detectors which learn multi-scale features or feature pyramids aggregated from different scale feature maps, SSFD only depends on a single-scale input image and a single-scale feature map to detect faces of various scales. In SSFD, transposed convolutions are leveraged to increase the resolution of feature maps with different strides, which can maintain adequate information for occluded and small faces. In addition, dilated convolutions are deployed to increase the receptive field size, which contributes to obtaining discriminative contextual information. SSFD, which is based on the VGG-16 network, outperforms the ResNet101-based Scale-Face as well as the VGG16-based HR on the WIDER FACE validation dataset.