.png){kind=icon}
返回
Optimizing offset-regression by relay point for bottom-up human pose estimation
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-023-05133-y.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
The regression of human pose from a central point of the person instance has garnered attention in bottom-up human pose estimation. While center-point regression approaches are notably efficient, their accuracy is compromised due to the challenges of predicting distant joints through center-to-joint offset. In this paper, a new geometrical point, named a relay point, is introduced to optimize the regression of long-range offset. With the relay point, we propose a relay pose representation (RPR), where the center-to-joint offset is factorized into two short-range offsets. This novel representation can reduce the impact of occlusion and the difficulty of directly regressing long-range offset. Toward an offset merging branch, the spatial information and semantic information of these two short-range offsets were merged to regress a more accurate offset. Moreover, in the grouping strategy, the initially predicted pose was optimized by a by-product of the network, relay point-to-joint offset, to further reduce the error introduced by long-range offset. Extensive experiments conducted on the MS-COCO dataset and the CrowdPose dataset demonstrate that the performance of our approach is comparable to that of current state-of-the-art bottom-up methods.
Keyword:
Bottom-up
Human pose estimation
Relay point
Guided spatial transform
期刊
IF:
3.5
论文数:
7.6K
被引数:
1.7W
机构
引用论文
.png){kind=icon}
查看更多内容