基于改进热图损失函数的目标6D姿态估计算法

Object 6D pose estimation algorithm based on improved heatmap loss function

针对传统热图回归使用的均方误差(MSE)损失函数训练热图回归网络的精度不高且训练缓慢的问题,本文提出了用于热图回归的损失函数Heatmap Wing Loss(HWing Loss)。该损失函数对于不同的像素值有不同的损失函数值,前景像素的损失函数梯度更大,可以使网络更加关注前景像素,使热图回归更加准确快速。同时根据热图分布特性,使用基于高斯分布的关键点推理方法减小热图推断关键点时的量化误差。以此两点为基础,构造新的基于关键点定位的单目标姿态估计的算法。实验结果表明,相比于使用MSE Loss的算法,使用HWing Loss的姿态估计算法有更高的ADD(-S)准确率,在LINEMOD数据集上达到了88.8%,性能优于近期其他的基于深度学习的姿态估计算法。本文算法在RTX3080 GPU上最快能以25 fps的速度运行,兼具速度与性能优势。

In view of the problem of low precision and slow training of heatmap regression network trained by mean square error(MSE)loss function used in traditional heatmap regression,the loss function Heatmap Wing Loss(HWing Loss)for heatmap regression is proposed in this thesis.In terms of different pixel values,the loss function has different loss function values,and the loss function gradient of foreground pixels is larger,which can make the network focus more on the foreground pixels and make the heatmap regression more accurate and faster.In line with the distribution characteristics of the heatmap,the keypoint inference method based on the Gaussian distribution is adopted in this thesis to reduce the quantization error when the heatmap infers the keypoints.By taking the two points as the basis,it constructs a new monocular pose estimation algorithm based on keypoint positioning.According to the experiments,in contrast with the algorithm using MSE Loss,the pose estimation algorithm using HWing Loss has a higher ADD(-S)accuracy rate,which reaches 88.8%on the LINEMOD dataset.Meanwhile,the performance is better than other recent pose estimation algorithms based on deep learning.The algorithm in this thesis can run at the fastest speed of 25 fps on RTX3080 GPU,in which the high speed and performance can be both embodied.