Lowlight Human Pose Estimation Using Depthwise Separable Convolution Curve Matching Networks

To improve the performance of human pose estimation in low luminance and to solve the problem of low data with annotations. This paper proposed a network framework LMPE-Net(Lowlight Matching Pose Estimate) for human pose estimation that combines separable convolution with higher-order curve matching, an unsupervised method for pixel-level higher-order curve matching is used to augment the lowlight image, the network parameters and complexity are further reduced by separable convolution instead of the standard convolution module. The COCO dataset and MPII dataset are used to train the human pose estimator in the experiments, in order to address the problem of the small amount of existing human pose data with annotated low light, the COCO and MPII human pose datasets are darkened, and the dataset is used to conduct a comparison experiment of low light pose estimation. Compared with similar state-of-the-art models, the experiments of the proposed DSLMPE framework show that the proposed lowlight image enhancement achieves 16.57 peak signal-to-noise ratio and 0.59 structural similarity and reduces the average absolute error by 0.8% in the experiments. And 92.3% PCK (Percentage of Correct Key points) is achieved in lowlight human pose estimation comparison experiments.