SCite‐HRNet: A self‐calibrating efficient network for pose estimation

Abstract In order to tackle the challenge of capturing long‐range spatial dependencies among joints, a novel self‐calibrated lightweight high‐resolution network (SCite‐HRNet), which is grounded on the lightweight high‐resolution network is introduced. A self‐calibrated segmentation convolution is first designed to extract and amalgamate contextual information across various scales, thereby addressing the issue of excessive computation engendered by stacked convolution kernels in conventional convolution methods. Next, a multi‐scale channel attention mechanism designed to extract structural information while filtering out unnecessary channel details is introduced. Ultimately, these two methodologies are incorporated into a multi‐scale information aggregation module and embed this module into the high‐resolution network. This allows the network to maintain exceptionally efficient computation while effectively managing information across various scales. Empirical results indicate that SCite‐HRNet achieves remarkable performance on both the COCO dataset and the challenging average precision (AP)‐10K dataset.