Importance relation guided spatio-temporal adaptive modeling framework for irregular 2D human pose estimation

Human pose estimation is an important task in computer vision and has been widely applied in many fields. Existing algorithms have achieved promising performance in regular motion scenarios, but estimating poses under irregular motions caused by body flipping, folding, fast motion, and occlusion remains challenging. To address this problem, we propose an Importance Relation Guided Spatio-Temporal Adaptive Modeling framework and introduce a self-built Irregular Motion Dataset mainly composed of high-intensity sports activities. In the spatial dimension, the framework incorporates a Multi-Scale Adaptive Topology Modeling module, which models fine-grained keypoint features, adaptively weights the adjacency matrix according to keypoint attributes, and enriches feature representation through multi-scale modeling. In the temporal dimension, a Global Relation Memory module separates historical features into short-term and long-term components, and selectively integrates useful temporal cues into the current frame according to their importance. This design helps reduce the influence of low-quality historical features caused by motion blur and occlusion. In addition, a Motion Trend Modeling module introduces keypoint-level motion trend cues to provide auxiliary directional information for continuous pose modeling. The self-built dataset provides a focused experimental benchmark for evaluating pose estimation methods under high-intensity irregular motions. Experiments on PoseTrack 2017, PoseTrack 2018, Penn Action, and the self-built dataset, together with ablation studies and visualizations, demonstrate the effectiveness of the proposed framework under the evaluated settings.