A hybrid classification-regression approach for 3D hand pose estimation using graph convolutional networks

作者：

Highlights：

• We introduce a two-stage graph-based approach that combines classification and regression to estimate an accurate 3D hand pose from a single RGB image.

• Exploiting spatial dependencies between neighboring hand joints to learn per-image relationship constraints.

• An adaptative nearest neighbor algorithm that learns a different number of joint relationships.

• Outperforming state-of-the-art on 2D/3D realistic and synthetic datasets and speeding up the inference time.

摘要

•We introduce a two-stage graph-based approach that combines classification and regression to estimate an accurate 3D hand pose from a single RGB image.•Exploiting spatial dependencies between neighboring hand joints to learn per-image relationship constraints.•An adaptative nearest neighbor algorithm that learns a different number of joint relationships.•Outperforming state-of-the-art on 2D/3D realistic and synthetic datasets and speeding up the inference time.

论文关键词：3D hand pose estimation,Graph convolutional networks,Classification,Multi-stage learning,Monocular RGB image

论文评审过程：Received 19 January 2021, Revised 26 June 2021, Accepted 14 October 2021, Available online 17 November 2021, Version of Record 25 November 2021.

论文官网地址：https://doi.org/10.1016/j.image.2021.116564