.png){kind=icon}
.png){kind=icon}
Return
RCENet: an efficient pose estimation network based on regression correction
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-024-01496-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Human pose estimation networks based on heatmap detection still have some problems including redundant computation, quantization errors, and non-differentiable heatmap decoding, which reduce the efficiency of these networks. To address these problems, we propose an efficient pose estimation network based on regression correction, i.e., RCENet. In the network, we design a partial convolution module based on coordinate weighting, and the module learns channel coordinate weights to compensate for the loss of global channel information in the partial convolution residual channels. The backbone network based on the module not only reduces channel redundancy but also improves network efficiency. As a common and important post-processing operation of keypoint extraction, heatmap decoding exists quantization errors that lead to the location deviation of keypoints, and the non-differentiable property of maximum probability computation causes the difficulty in embedding heatmap decoding into end-to-end learning of pose estimation networks. Thus, we construct a novel regression correction (RC) module. In RC, a non-global dependent integral regression is given for reducing the bias of expectation computation in integral regression, and an extra one-dimensional bias compensation is designed for further correcting the bias of integral regression. RC without any post-processing operation can be directly fused into the end-to-end learning of keypoint coordinates and can correct the quantization errors. Experimental results show that our RCENet competes well on challenging pose estimation datasets.
Keywords:
Pose estimation
High-resolution network
Integral regression
Coordinate weighting
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
No organization information available