基于<b>DDPG</b>的三维重建模糊概率点推理

李雷; 徐浩; 吴素萍

doi:10.16383/j.aas.c200543

基于DDPG的三维重建模糊概率点推理

doi: 10.16383/j.aas.c200543

李雷^1,,
徐浩^1,,
吴素萍^1,

1.
宁夏大学信息工程学院银川 750021

基金项目: 国家自然科学基金(62062056, 61662059)资助

详细信息

作者简介:
李雷：宁夏大学信息工程学院硕士研究生. 主要研究方向为三维物体重建, 人脸重建以及关键点对齐, 图像处理和计算机视觉与模式识别. E-mail: lliicnxu@163.com

徐浩：宁夏大学信息工程学院硕士研究生. 主要研究方向为计算机视觉和三维人体姿态估计. E-mail: hao_xu321@163.com

吴素萍：宁夏大学信息工程学院教授. 主要研究方向为三维重建, 计算机视觉, 模式识别, 并行分布处理与大数据. 本文通信作者. E-mail: pswuu@nxu.edu.cn

计量
- 文章访问数: 1532
- HTML全文浏览量: 454
- PDF下载量: 196
- 被引次数: 0
出版历程
- 收稿日期: 2020-07-13
- 修回日期: 2020-12-05
- 网络出版日期: 2021-03-02
- 刊出日期: 2022-04-13

Fuzzy Probability Points Reasoning for 3D Reconstruction Via Deep Deterministic Policy Gradient

LI Lei^1
,,
XU Hao^1
,,
WU Su-Ping^1
,

1.
School of Information Engineering, Ningxia University, Yinchuan 750021

Funds: Supported by National Natural Science Foundation of China (62062056, 61662059)

More Information

Author Bio:
LI Lei　Master student at the School of Information Engineering, Ningxia University. His research interest covers 3D object reconstruction, face reconstruction and landmark alignment, image processing, computer vision and pattern recognition

XU Hao　Master student at the School of Information Engineering, Ningxia University. His research interest covers computer vision and 3D human pose estimation

WU Su-Ping　Professor at the School of Information Engineerring, Ningxia University. Her research interest covers 3D reconstruction, computer vision, pattern recognition, parallel distributed processing and big data. Corresponding author of this paper

摘要

摘要: 单视图物体三维重建是一个长期存在的具有挑战性的问题. 为了解决具有复杂拓扑结构的物体以及一些高保真度的表面细节信息仍然难以准确进行恢复的问题, 本文提出了一种基于深度强化学习算法深度确定性策略梯度 (Deep deterministic policy gradient, DDPG)的方法对三维重建中模糊概率点进行再推理, 实现了具有高保真和丰富细节的单视图三维重建. 本文的方法是端到端的, 包括以下四个部分: 拟合物体三维形状的动态分支代偿网络的学习过程, 聚合模糊概率点周围点的邻域路由机制, 注意力机制引导的信息聚合和基于深度强化学习算法的模糊概率调整. 本文在公开的大规模三维形状数据集上进行了大量的实验证明了本文方法的正确性和有效性. 本文提出的方法结合了强化学习和深度学习, 聚合了模糊概率点周围的局部信息和图像全局信息, 从而有效地提升了模型对复杂拓扑结构和高保真度的细节信息的重建能力.
- 三维重建 /
- 强化学习 /
- 深度学习 /
- 注意力机制 /
- 信息聚合
Abstract: 3D object reconstruction from a single-view image is a long-standing challenging problem. In order to address the difficulty of accurately predicting the objects of complex topologies and some high-fidelity surface details, we propose a new method based on DDPG (Deep deterministic policy gradient) to reason the fuzzy probability points in 3D reconstruction and achieve high-quality detail-rich reconstruction result of single-view image. Our method is end-to-end and includes four parts: the dynamic branch compensation network learning process to fit the 3D shape of objects, the neighborhood routing mechanism to aggregate the points around the fuzzy probability points, the attention guidance mechanism to aggregate the information, and the deep reinforcement learning algorithm to perform probabilistic reasoning. Extensive experiments on a large-scale public 3D shape dataset demonstrate the validity and efficiency of our method. Our method combines reinforcement learning and deep learning, aggregates local information around the fuzzy probability points and global information of the image, and effectively improves the model＇s ability to reconstruct complex topologies and high-fidelity details.
- 3D reconstruction /
- reinforcement learning /
- deep learning /
- attention mechanism /
- information aggregation

HTML全文

图 1 基于深度学习的单视图三维重建中三种表示形状

Fig. 1 Three representation shapes for single-view 3D reconstruction based on deep learning

下载: 全尺寸图片幻灯片

图 2 本文方法和DISN方法在真实图像上的单视图重建结果

Fig. 2 Single image reconstruction using a DISN, and our method on real images

下载: 全尺寸图片幻灯片

图 3 MNGD框架的整体流程图

Fig. 3 The workflow of the proposed MNGD framework

下载: 全尺寸图片幻灯片

图 4 动态分支代偿网络框架图

Fig. 4 The framework of the dynamic branch compensation network

下载: 全尺寸图片幻灯片

图 5 邻域路由过程

Fig. 5 The whole process of neighbor routing

下载: 全尺寸图片幻灯片

图 6 聚合特征时的注意力机制

Fig. 6 Attention mechanism when features are aggregated

下载: 全尺寸图片幻灯片

图 7 卷积可视化与网格生成过程

Fig. 7 Convolution visualization and mesh generation process

下载: 全尺寸图片幻灯片

图 8 ShapeNet数据集上的定性结果

Fig. 8 Qualitative results on the ShapeNet dataset

下载: 全尺寸图片幻灯片

图 9 Online Products dataset的定性结果

Fig. 9 Qualitative results on Online Products dataset

下载: 全尺寸图片幻灯片

图 10 消融实验的定性结果

Fig. 10 Qualitative results of ablation study

下载: 全尺寸图片幻灯片

图 11 MNGD随机调整100张图片中模糊概率点的结果

Fig. 11 The result of MNGD adjusting the fuzzy probability points in 100 random images

下载: 全尺寸图片幻灯片

图 12 ShapeNet上所有类别的定性结果

Fig. 12 Qualitative results on ShapeNet of all categories

下载: 全尺寸图片幻灯片

图 13 单视图三维重建中具有挑战性案例

Fig. 13 Challenging cases in single-view 3D reconstruction

下载: 全尺寸图片幻灯片

表 1 本文的方法在ShapeNet数据集上与最先进方法的交并比(IoU)的定量比较

Table 1 The quantitative comparison of our method with the state-of-the-art methods for IoU on ShapeNet dataset

类别\方法	3D-R2N2	Pix2Mesh	AtlasNet	ONet	Our
Airplane	0.426	0.420	—	0.571	0.592
Bench	0.373	0.323	—	0.485	0.503
cabinet	0.667	0.664	—	0.733	0.757
Car	0.661	0.552	—	0.737	0.755
Chair	0.439	0.396	—	0.501	0.542
Display	0.440	0.490	—	0.471	0.548
Lamp	0.281	0.323	—	0.371	0.409
Loudspeaker	0.611	0.599	—	0.647	0.672
Rifle	0.375	0.402	—	0.474	0.500
Sofa	0.626	0.613	—	0.680	0.701
Table	0.420	0.395	—	0.506	0.547
Telephone	0.611	0.661	—	0.720	0.763
Vessel	0.482	0.397	—	0.530	0.569
Mean	0.493	0.480	—	0.571	0.605

下载: 导出CSV

表 2 本文的方法在ShapeNet数据集上与最先进方法法线一致性(NC)的定量比较

Table 2 The quantitative comparison of our method with the state-of-the-art methods for NC on ShapeNet dataset

类别\方法	3D-R2N2	Pix2Mesh	AtlasNet	ONet	Our
Airplane	0.629	0.759	0.836	0.840	0.847
Bench	0.678	0.732	0.779	0.813	0.818
Cabinet	0.782	0.834	0.850	0.879	0.887
Car	0.714	0.756	0.836	0.852	0.855
Chair	0.663	0.746	0.791	0.823	0.835
Display	0.720	0.830	0.858	0.854	0.871
Lamp	0.560	0.666	0.694	0.731	0.751
Loudspeaker	0.711	0.782	0.825	0.832	0.845
Rifle	0.670	0.718	0.725	0.766	0.781
Sofa	0.731	0.820	0.840	0.863	0.872
Table	0.732	0.784	0.832	0.858	0.864
Telephone	0.817	0.907	0.923	0.935	0.938
Vessel	0.629	0.699	0.756	0.794	0.801
Mean	0.695	0.772	0.811	0.834	0.844

下载: 导出CSV

表 3 本文的方法在ShapeNet数据集上与最先进方法倒角距离 (CD)的定量比较

Table 3 The quantitative comparison of our method with the state-of-the-art methods for CD on ShapeNet dataset

类别\方法	3D-R2N2	Pix2Mesh	AtlasNet	ONet	Our
Airplane	0.227	0.187	0.104	0.147	0.130
Bench	0.194	0.201	0.138	0.155	0.149
Cabinet	0.217	0.196	0.175	0.167	0.146
Car	0.213	0.180	0.141	0.159	0.144
Chair	0.270	0.265	0.209	0.228	0.200
Display	0.314	0.239	0.198	0.278	0.220
Lamp	0.778	0.308	0.305	0.479	0.364
Loudspeaker	0.318	0.285	0.245	0.300	0.263
Rifle	0.183	0.164	0.115	0.141	0.130
Sofa	0.229	0.212	0.177	0.194	0.179
Table	0.239	0.218	0.190	0.189	0.170
Telephone	0.195	0.149	0.128	0.140	0.121
Vessel	0.238	0.212	0.151	0.218	0.189
Mean	0.278	0.216	0.175	0.215	0.185

下载: 导出CSV

表 4 消融实验

Table 4 Ablation study

模型\指标	IoU	NC	CD
FM w/o DR, MB	0.593	0.840	0.194
FM w/o MB	0.599	0.839	0.194
FM	0.605	0.844	0.185

下载: 导出CSV

参考文献(44)

[1]	陈加, 张玉麒, 宋鹏, 魏艳涛, 王煜. 深度学习在基于单幅图像的物体三维重建中的应用. 自动化学报, 2019, 45(4): 657-668 Chen Jia, Zhang Yu-Qi, Song Peng, Wei Yan-Tao, Wang Yu. Application of deep learning to 3D object reconstruction from a single image. Acta Automatica Sinica, 2019, 45(4): 657-668
[2]	郑太雄, 黄帅, 李永福, 冯明驰. 基于视觉的三维重建关键技术研究综述. 自动化学报, 2020, 46(4): 631-652 Zheng Tai-Xiong, Huang Shuai, Li Yong-Fu, Feng Ming-Chi. Key techniques for vision based 3D reconstruction: A review. Acta Automatica Sinica, 2020, 46(4): 631-652
[3]	薛俊诗, 易辉, 吴止锾, 陈向宁. 一种基于场景图分割的混合式多视图三维重建方法. 自动化学报, 2020, 46(4): 782-795 Xue Jun-Shi, Yi Hui, Wu Zhi-Huan, Chen Xiang-Ning. A hybrid multi-view 3D reconstruction method based on scene graph partition. Acta Automatica Sinica, 2020, 46(4): 782-795
[4]	Wu J J, Zhang C K, Xue T F, Freeman W T, Tenenbaum J B. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In: Proceedings of the 30th International Conference on Neural Information Processing Systems. Barcelona, Spain: Curran Associates, Inc., 2016. 82−90
[5]	Choy C B, Xu D F, Gwak J Y, Chen K, Savarese S. 3D-R2N2: A unified approach for single and multi-view 3D object reconstruction. In: Proceedings of the 14th European Conference on Computer Vision. Amsterdam, The Netherlands: Springer, 2016. 628−644
[6]	Yao Y, Luo Z X, Li S W, Fang T, Quan L. MVSNet: Depth inference for unstructured multi-view stereo. In: Proceedings of the 15th European Conference on Computer Vision. Munich, Germany: Springer, 2018. 785−801
[7]	Wu J J, Wang Y F, Xue T F, Sun X Y, Freeman W T, Tenenbaum J B. MarrNet: 3D shape reconstruction via 2.5D sketches. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates, Inc., 2017. 540−550
[8]	Fan H Q, Su H, Guibas L. A point set generation network for 3D object reconstruction from a single image. In: Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, USA: IEEE, 2017. 2463−2471
[9]	Wang N Y, Zhang Y D, Li Z W, Fu Y W, Liu W, Jiang Y G. Pixel2Mesh: Generating 3D mesh models from single RGB images. In: Proceedings of the 15th European Conference on Computer Vision. Munich, Germany: Springer, 2018. 55−71
[10]	Scarselli F, Gori M, Tsoi A C, Hagenbuchner M, Monfardini G. The graph neural network model. IEEE Transactions on Neural Networks, 2009, 20(1): 61-80 doi: 10.1109/TNN.2008.2005605
[11]	Rumelhart D E, Hinton G E, Williams R J. Learning representations by back-propagating errors. Nature, 1986, 323(6088): 533-536 doi: 10.1038/323533a0
[12]	Roth S, Richter S R. Matryoshka networks: Predicting 3D geometry via nested shape layers. In: Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 1936−1944
[13]	Wu J J, Zhang C K, Zhang X M, Zhang Z T, Freeman W T, Tenenbaum J B. Learning shape priors for single-view 3D completion and reconstruction. In: Proceedings of the 15th European Conference on Computer Vision. Munich, Germany: Springer, 2018. 673−691
[14]	Groueix T, Fisher M, Kim V G, Russell B C, Aubry M. A Papier-Mache approach to learning 3D surface generation. In: Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 216−224
[15]	Kanazawa A, Black M J, Jacobs D W, Malik J. End-to-end recovery of human shape and pose. In: Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 7122−7131
[16]	Kong C, Lin C H, Lucey S. Using locally corresponding CAD models for dense 3D reconstructions from a single image. In: Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, USA: IEEE, 2017. 5603−5611
[17]	Mescheder L, Oechsle M, Niemeyer M, Nowozin S, Geiger A. Occupancy networks: Learning 3D reconstruction in function space. In: Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE, 2019. 4455−4465
[18]	Lillicrap T P, Hunt J J, Pritzel A, Heess N, Erez T, Tassa Y, et al. Continuous control with deep reinforcement learning [Online], available: https: //arxiv.org/abs/1509. 02971, July 5, 2019
[19]	Li D, Chen Q F. Dynamic hierarchical mimicking towards consistent optimization objectives. In: Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle, USA: IEEE, 2020. 7639−7648
[20]	Chang A X, Funkhouser T, Guibas L, et al.Shapenet:An information-rich 3d model repository [Online], available: https: //arxiv.org/abs/1512. 03012, December 9, 2015
[21]	Durou J D, Falcone M, Sagona M. Numerical methods for shape-from-shading: A new survey with benchmarks. Computer Vision and Image Understanding, 2008, 109(1): 22-43 doi: 10.1016/j.cviu.2007.09.003
[22]	Zhang R, Tsai P S, Cryer J E, Shah M. Shape-from-shading: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1999, 21(8): 690-706 doi: 10.1109/34.784284
[23]	Goodfellow I J, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Generative adversarial nets. In: Proceedings of the 27th International Conference on Neural Information Processing Systems. Montreal, Canada: MIT Press, 2014. 2672−2680
[24]	Kingma D P, Welling M. Auto-encoding variational bayes [Online], available: https: //arxiv. org/abs/1312. 6114, May 1, 2014
[25]	Kar A, Hane C, Malik J. Learning a multi-view stereo machine. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates, Inc., 2017. 364−375
[26]	Tatarchenko M, Dosovitskiy A, Brox T. Octree generating networks: Efficient convolutional architectures for high-resolution 3D outputs. In: Proceedings of the 2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy: IEEE, 2017. 2107−2115
[27]	Wang W Y, Ceylan D, Mech R, Neumann U. 3DN: 3D deformation network. In: Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE, 2019. 1038−1046
[28]	Bernardini F, Mittleman J, Rushmeier H, Silva C, Taubin G. The ball-pivoting algorithm for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 1999, 5(4): 349-359 doi: 10.1109/2945.817351
[29]	Kazhdan M, Hoppe H. Screened poisson surface reconstruction. ACM Transactions on Graphics, 2013, 32(3): Article No. 29
[30]	Calakli F, Taubin G. SSD: Smooth signed distance surface reconstruction. Computer Graphics Forum, 2011, 30(7): 1993-2002 doi: 10.1111/j.1467-8659.2011.02058.x
[31]	Chen Z Q, Zhang H. Learning implicit fields for generative shape modeling. In: Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, USA: IEEE, 2019. 5932−5941
[32]	Wang W Y, Xu Q G, Ceylan D, Mech R, Neumann U. DISN: Deep implicit surface network for high-quality single-view 3D reconstruction. In: Proceedings of the 33rd International Conference on Neural Information Processing Systems. Vancouver, Canada: Curran Associates, Inc., 2019. Article No. 45
[33]	Wang Q L, Wu B G, Zhu P F, Li P H, Zuo W M, Hu Q H. ECA-Net: Efficient channel attention for deep convolutional neural networks. In: Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle, USA: IEEE, 2020. 11531−11539
[34]	Selvaraju R R, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the 2017 IEEE International Conference on Computer Vision (ICCV). Venice, Italy: IEEE, 2017. 618−626
[35]	Garland M, Heckbert P S. Simplifying surfaces with color and texture using quadric error metrics. In: Proceedings of the 1998 Visualization′ 98 (Cat. No.98CB362-76). Research Triangle Park, USA: IEEE, 1998. 263−269
[36]	Lorensen W E, Cline H E. Marching cubes: A high resolution 3D surface construction algorithm. ACM SIGGRAPH Computer Graphics, 1987, 21(4): 163-169 doi: 10.1145/37402.37422
[37]	Drucker H, Le Cun Y. Improving generalization performance using double backpropagation. IEEE Transactions on Neural Networks, 1992, 3(6): 991-997 doi: 10.1109/72.165600
[38]	Oh Song H, Xiang Y, Jegelka S, Savarese S. Deep metric learning via lifted structured feature embedding. In: Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE, 2016. 4004−4012
[39]	Stutz D, Geiger A. Learning 3D shape completion from laser scan data with weak supervision. In: Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 1955−1964
[40]	de Vries H, Strub F, Mary J, Larochelle H, Pietquin O, Courville A C. Modulating early visual processing by language. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates, Inc., 2017. 6597−6607
[41]	He K M, Zhang X Y, Ren S Q, Sun J. Deep residual learning for image recognition. In: Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, USA: IEEE, 2016. 770−778
[42]	Kingma D P, Ba J. Adam: A method for stochastic optimization [Online], available: https: //arxiv. org/abs/1412. 6980, January 30, 2017
[43]	Zhu C C, Liu H, Yu Z H, Sun, X H. Towards Omni-supervised face alignment for large scale unlabeled videos. In: Proceedings of the 34th AAAI Conference on Artificial Intelligence. New York, USA: AAAI, 2020. 13090−13097
[44]	Zhu C C, Li X Q, Li J D, Ding G T, Tong W Q. Spatial-temporal knowledge integration: Robust self-supervised facial landmark tracking. In: Proceedings of the 28th ACM International Conference on Multimedia. Lisboa, Portugal: ACM, 2020. 4135−4143