基于全局分组策略的对抗训练在人体轮廓点回归中的应用模型

基于深度学习对人体姿态或人体骨骼的关键点进行预测是近期计算机视觉方向上的一个热门领域,取得了大量的研究成果。然而,在部分应用场景中(如人体特定部位瘦身、衣装更换等),传统的人体姿态或骨骼关键点的预测并不能完全满足需求,其他技术(如语义分割)在自遮挡情况下也无法准确描述物体轮廓线。针对人体轮廓点的预测能力进行研究,分析了现有技术应用在轮廓点回归上的问题与不足,提出全局和分组的策略对轮廓点进行预测,并引入对抗训练机制以提升系统整体性能。基于提出的方法在自建数据集上进行了一系列实验,实验结果证明,该方法相较于传统的骨骼回归模型在精准度和计算量上都有较大的优势。

Human outline keypoints detecting via global and grouping strategy based on GAN

Pose estimation based on deep learning is a hot topic in recent years.Different from it,the outline keypoints detecting task has not yet long been researched sufficiently in computer vision field.Body’s outline cannot be recovered directly with joint keypoints or skeleton only,even with the aid of semantic segmentation.This paper aimed at detecting points of human’s outline with ordered keypoints.By analyzing the characters of the task,it put forward global and grouping strategy and processed with adversarial training to overcome the traditional methods’defects.Experimental results over self-build dataset show that the proposed method has excellent state-of-the-art performance compared with traditional pose estimation models.