Multi Person Pose Estimation with Attention

The existing pose estimation networks extract confidence maps and Part Affinity Fields (PAFs) from all pixels in images. However, we found that the recognition rate can be decreased due to the background. This limitation arises from the process to extract feature maps from the image. To overcome this limitation, we adopt the attention block in front of the backbone network. To show that our proposed network can reduce the error and increase the precision, we compared PCKh and mAP with the existing pose estimation network.