In the fields of advertising and fashion, the pose of a professional model plays a pivotal role in highlighting the distinct features and aesthetics of apparel. Models rely on observing images, videos, or classroom teaching to train their body poses. This process often faces challenges in grasping the nuances of movements and lacking digital and standardized feedback, thus hindering the development of correct pose habits. Additionally, the clothing, hairstyles, styling, and actions exhibited by models are diverse and complex, making it difficult for existing human pose estimation methods to accurately capture and locate the display posture of models. Addressing these limitations, this paper constructs a dataset of 12,284 images, encompassing professional models in various poses, and defines 20 key points across 11 critical pose areas. Based on this, an improved stacked network model is proposed to accurately locate keypoints of models. The model consists of a backbone network composed of multiple stacked V modules and integrates attention mechanisms and depth-adaptive intermediate supervision to enhance the model's expressive capability, training speed, and localization precision. Comprehensive experiments conducted on the constructed model pose dataset demonstrate the superiority of the proposed algorithm, and further ablation study analysis verifies the effectiveness of each module within the proposed method.