Year: 2025
Type: article
Abstract: Aiming at the insufficient localization accuracy of traditional algorithms due to complex texture interference, diverse fabric deformations and sensitivity to small size errors in garment size detection, this paper proposes an improved HRNet-cloth key point detection model. By introducing full-dimensional dynamic convolution (ODConv) in the HRNet backbone network, we enhance the feature adaptation ability of the network to nonlinear deformation such as garment folds and draping, and effectively reduce the key point coordinate offset error; we design the EMA cross-dimensional attention mechanism module, fusing the channel and spatial dimensional feature responses to improve the localization robustness of the neckline, sleeve holes, and other detail regions; for the sub-pixel level regression requirements, we Construct an adaptive focus loss function to optimize the heat map peak distribution by dynamically adjusting the weights of difficult samples. Experiments show that the PR of HRNet-cloth on the self-built dataset ClothData reaches 100%, which is 11.6% higher than that of the benchmark model, and the absolute measurement error (AKE) of the dimensions is stabilized within ±1cm.
DOI: 10.54097/9wm2g323
Authors: Zhuohui Li, Yanfang Fu
Institution: Xi'an Technological University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4409133518