Development of Face and Landmark Detection using EfficientNetV2 and BiFPN

Abstract

In this paper, we develop improved face and landmark detection algorithm using EfficientNetV2 as backbone and BiFPN as multi-scale feature extractor. EfficientNetV2 are a new family of convolutional networks that have faster training speed and better parameter efficiency than previous models. BiFPN is a new multi-scale feature extractor that allows easy and fast multi-scale fusion. To evaluate the performance of the proposed face and landmark detection algorithm, we trained and tested on WIDER FACE dataset using PyTorch framework in NVIDIA Titan RTX GPU. In the experiments, we show the experimental results for comparing the detection accuracy and efficiency of the proposed network to MobileNetV1 0.25 and Resnet50 networks. The experimental results show that the proposed algorithm is accurate and efficient than previous works.