摘要
在体素化输入的三维卷积基础上,通过引入高分辨率保持模块,提出了一种基于三维卷积的高分辨率保持网络.该网络以体素化的深度图为输入,进行三维卷积.不同于先前大多数从低分辨率特征中恢复高分辨率特征的方法,构建的网络引入不同分辨率子网络并行的结构,在处理低分辨率特征图的同时保持高分辨率特征图,从高分辨率子网络卷积得出每个关节点在3D体素中的分布概率,最终计算出每个关节点的三维空间坐标.实验表明:该算法相较于先前的基于沙漏模型的三维卷积网络能更准确地进行关节点估计.
A high resolution deep neuron network based on three dimensional(3 D) convolution was proposed by introducing high resolution maintain modules.This network took voxelized depth map as input and performed three-dimensional convolution.Different from most existing methods recovering high-resolution representations from low-resolution representations,the network was constructed by introducing parallel structures of sub-networks with different resolutions,which can process low-resolution feature maps while maintaining high-resolution feature maps.The network results could be obtained from the final high-resolution representations in the form of per-voxel likelihood for each keypoint,from which 3 D location of each key point could be obtained.Experiments show that the proposed network can generate more accurate keypoints’ 3 D locations compared with form networks based on hourglass module.
作者
桑农
李默然
SANG Nong;LI Moran(Key Laboratory of Ministry of Education for Image Processing and Intelligent Control,School of Artitical Intelligence and Automation,Huazhong University of Science and Technology,Wuhan 430074,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2020年第1期1-6,共6页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金面上项目(61871435)
关键词
人手姿态估计
单一深度图
体素化
3D卷积
高分辨率网络
human hand pose estimation
single depth map
voxelize
3D convolution
high resolution network
