EG-HRNet: An Efficient High-Resolution Network Using Ghost-Modules for Human Pose Estimation

As an essential task in predicting a human’s behavior, human pose estimation (HPE) plays a very important role in many real-time applications. However, existing HPE methods are still too large which severely prevents them to be used in resource-sensitive applications. In this paper, aiming to a significant reduction in computation complexity, we propose an efficient high-resolution HPE network using ghost-modules (EG-HRNet). Based on the HRNet architecture, the new EG-HRNet uses modified shuffle blocks as the inner blocks to replace the residual blocks. Meanwhile, we use the lightweight ghost bottleneck for a more efficient feature extraction and use the ghost modules in the fusion layers to replace the costly 1x1 point-wise convolutions. Finally, we use the distribution-aware coordinate representation of the keypoints to acquire more accurate heatmaps of the input images. The experimental results on the COCO keypoint detection dataset show that the new efficient EG-HRNet model has successfully reached a tender balance between the processing speed and the estimation accuracy.