输入一张人体图像,输出人体参数化模型SMPL的pose和shape估计结果,已经对应的vertices,3D关键点。
SMPL模型
pose 维度(24, 3),包含1个global rotation与23个local joints rotation,用三维的旋转向量表示
shape 维度(1, 10)
vertices 维度(6890, 3),人体Mesh的6890个顶点,faces为SMPL模型的标准faces。
模型描述
模型来源CVPR 2023工作 Towards Stable Human Pose Estimation via Cross-View Fusion and Foot Stabilization,由三部分组成:
- ViT作为backbone用于提取图像特征
- Cross-View Fusion module用于预测front-view 2D heatmaps,同时基于front-view 2D heatmaps与图像特征进一步得到side-view与top-view的2D heatmaps。
- RKTD用于辅助脚部姿态预测。(本模型中不包含此模块)
推理代码范例
引用
@InProceedings{Zhuo_2023_CVPR,
author = {Zhuo, Li{\textquoteright}an and Cao, Jian and Wang, Qi and Zhang, Bang and Bo, Liefeng},
title = {Towards Stable Human Pose Estimation via Cross-View Fusion and Foot Stabilization},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2023},
pages = {650-659}
}
评论