摘 要：基于RGBD 的6D 位姿估计方法的一个关键问题是如何进行彩色特征信息和深度特征信息的融合。先前的工作采用密集融合的方法，主要关注的是局部特征和全连接层提取的全局特征，忽略了远距离像素间的位置依赖关系。文章提出通过捕获像素间的位置关系，并将其与彩色特征图和几何特征图进行密集融合，最后逐像素预测物体的6D 位姿。实验结果表明，该文的方法相比其他方法在YCB-Video 数据集上获得更优的结果。
关键词：6D 位姿估计；弱纹理；RGB-D；密集融合
中图分类号：TP751 文献标识码：A 文章编号：2096-4706（2020）22-0016-04
6D Pose Estimation Method Based on Position Dependent Dense Fusion
HUANG Rongbin
（Guangdong University of Technology，Guangzhou 510006，China）
Abstract：One of the key problems of the 6D pose estimation method based on RGBD is how to fuse the color feature information and depth feature information. Previous work used dense fusion method，mainly focused on local features and global features extracted from fully connected layer，ignoring the position dependence between remote pixels. The article proposes that by capturing the positional relationship between pixels and intensively fusing it with the color feature map and geometric feature map，the 6D pose of the object is predicted pixel by pixel. Experimental results show that the proposed method achieves better results than other methods on YCB-Video dataset.
Keywords：6D pose estimation；weak texture；RGB-D；dense fusion
参考文献：
[1] LUCA V，VINCENT L，PASCAL F. Stable real-time 3Dtracking using online and offline information [J].IEEE transactions onpattern analysis and machine intelligence，2004，26（10）：1385-1391.
[2] LOWE D G. Object recognition from local scale-invariantfeatures [C]//Proceedings of the Seventh IEEE International Conferenceon Computer Vision.Kerkyra：IEEE，1999：1150-1157.
[3] PENG S D，LIU Y，HUANG Q X，et al. PVNet：PixelwiseVoting Network for 6DoF Object Pose Estimation [C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition（CVPR）.Long Beach：IEEE，2019：4556-4565.
[4] KEHL W，MANHARDT F，TOMBARI F，et al. SSD-6D：Making RGB-Based 3D Detection and 6D Pose Estimation Great Again [C]//2017 IEEE International Conference on Computer Vision （ICCV）.Venice：IEEE，2017：1530-1538.
[5] WANG C，XU D F，ZHU Y K，et al. DenseFusion：6DObject Pose Estimation by Iterative Dense Fusion [C]//2019 IEEE/CVFConference on Computer Vision and Pattern Recognition （CVPR）.Long Beach：IEEE，2019：3338-3347.
[6] XU D F，ANGUELOV D，JAIN A. PointFusion：DeepSensor Fusion for 3D Bounding Box Estimation [C]//2018 IEEE/CVFConference on Computer Vision and Pattern Recognition.Salt LakeCity：IEEE，2018：244-253.
[7] WANG X L，GIRSHICK R，GUPTA A，et al. Non-localNeural Networks [C]//2018 IEEE/CVF Conference on Computer Visionand Pattern Recognition.Salt Lake City：IEEE，2018：7794-7803.
[8] X I A N G Y，S C H M I D T T，NARAYA N A N V，e t a l .PoseCNN：A Convolutional Neural Network for 6D Object PoseEstimation in Cluttered Scenes [J/OL].arXiv：1711.00199 [cs.CV].（2017-11-01）.https://arxiv.org/abs/1711.00199.
[9] QI C R，SU H，MO K C，et al. PointNet：Deep Learningon Point Sets for 3D Classification and Segmentation [C]//2017 IEEEConference on Computer Vision and Pattern Recognition（CVPR）.Honolulu：IEEE，2017：77-85.
作者简介：黄榕彬（1995—），男，汉族，广东揭阳人，硕士研究生在读，研究方向：6D 位姿估计。