为解决单目穿衣人体在复杂姿态下手部形状重建存在遮挡和缺失的失真问题,提出了一种结合ECON与MANO手部模型,实现高效穿衣人体的手部精细化重建方法H-ECON(hand-focused explicit clothed humans obtained from normials).具体而言,该方法首先以类型无关的手部检测器聚焦手部区域并进行翻转和裁剪;然后,引入注意力机制用于增强对手部区域的感知能力,空洞螺旋卷积则更好地捕捉手部不同尺度的特征;最后,独特的融合模块确保了手部重建与整身模型的融合效果.在FreiHAND和HanCo公开数据集上与其他方法的定量定性对比结果表明了 H-ECON的有效性,其独立手部模块明显优于ECON中的替代手部模块.H-ECON实现了对人体手部几何和姿态变化的精确描述,进一步缩小了 2D图像生成到3D人体网格之间的差距.
Abstract
To solve the problem of occlusion and missing distortion in hand shape reconstruction of monocular clothed human body under complex posture,this paper proposed an efficient hand refinement reconstruction method H-ECON,which com-bined ECON and MANO to achieve efficient hand refinement reconstruction in the clothed body.Specifically,the method first-ly focused on the hand region with a type-independent hand detector and performed flipping and cropping.Then,the method enhanced the perception ability of the hand area through attention mechanism,and dilated spiral convolution can better capture the features of the hand at different scales.Finally,the unique fusion module ensured effective integration between the hand reconstruction and the entire body model.Quantitative and qualitative comparisons with other methods on the FreiHAND and HanCo publicly available datasets demonstrate the effectiveness of H-ECON,and its standalone hand module is significantly superior to the alternative hand module in ECON.H-ECON enables accurate descriptions of the geometry and pose changes of human hands,further narrowing the gap between 2D image generation and 3D human mesh.
关键词
手部重建/穿衣人体/注意力机制/空洞螺旋卷积/深度几何学习
Key words
hand reconstruction/clothed human/attention mechanism/dilated spiral convolution/deep geometry learning
维普
万方数据