A comprehensive framework for occluded human pose estimation

Occlusion presents a significant challenge in human pose estimation. Thechallenges posed by occlusion can be attributed to the following factors: 1)Data: The collection and annotation of occluded human pose samples arerelatively challenging. 2) Feature: Occlusion can cause feature confusion dueto the high similarity between the target person and interfering individuals.3) Inference: Robust inference becomes challenging due to the loss of completebody structural information. The existing methods designed for occluded humanpose estimation usually focus on addressing only one of these factors. In thispaper, we propose a comprehensive framework DAG (Data, Attention, Graph) toaddress the performance degradation caused by occlusion. Specifically, weintroduce the mask joints with instance paste data augmentation technique tosimulate occlusion scenarios. Additionally, an Adaptive DiscriminativeAttention Module (ADAM) is proposed to effectively enhance the features oftarget individuals. Furthermore, we present the Feature-Guided Multi-Hop GCN(FGMP-GCN) to fully explore the prior knowledge of body structure and improvepose estimation results. Through extensive experiments conducted on threebenchmark datasets for occluded human pose estimation, we demonstrate that theproposed method outperforms existing methods. Code and data will be publiclyavailable.