On the Importance of Accurate Geometry Data for Dense 3D Vision Tasks

被引:8
|
作者
Jung, HyunJun [1 ]
Ruhkamp, Patrick [1 ,2 ]
Zhai, Guangyao [1 ]
Brasch, Nikolas [1 ]
Li, Yitong [1 ]
Verdie, Yannick [1 ,3 ]
Song, Jifei [3 ]
Zhou, Yiren [3 ]
Armagan, Anil [3 ]
Ilic, Slobodan [1 ,4 ]
Leonardis, Ales [3 ]
Navab, Nassir [1 ]
Busam, Benjamin [1 ,2 ]
机构
[1] Tech Univ Munich, Munich, Germany
[2] Dwe Ai, Munich, Germany
[3] Huawei Noahs Ark Lab, Montreal, PQ, Canada
[4] Siemens AG, Munich, Germany
来源
2023 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, CVPR | 2023年
关键词
D O I
10.1109/CVPR52729.2023.00082
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Learning-based methods to solve dense 3D vision problems typically train on 3D sensor data. The respectively used principle of measuring distances provides advantages and drawbacks. These are typically not compared nor discussed in the literature due to a lack of multi-modal datasets. Texture-less regions are problematic for structure from motion and stereo, reflective material poses issues for active sensing, and distances for translucent objects are intricate to measure with existing hardware. Training on inaccurate or corrupt data induces model bias and hampers generalisation capabilities. These effects remain unnoticed if the sensor measurement is considered as ground truth during the evaluation. This paper investigates the effect of sensor errors for the dense 3D vision tasks of depth estimation and reconstruction. We rigorously show the significant impact of sensor characteristics on the learned predictions and notice generalisation issues arising from various technologies in everyday household environments. For evaluation, we introduce a carefully designed dataset1 comprising measurements from commodity sensors, namely D-ToF, I-ToF, passive/active stereo, and monocular RGB+P. Our study quantifies the considerable sensor noise impact and paves the way to improved dense vision estimates and targeted data fusion.
引用
收藏
页码:780 / 791
页数:12
相关论文
共 50 条
  • [31] An autonomous active vision system for complete and accurate 3D scene reconstruction
    Marchand, É
    Chaumette, F
    INTERNATIONAL JOURNAL OF COMPUTER VISION, 1999, 32 (03) : 171 - 194
  • [32] An Autonomous Active Vision System for Complete and Accurate 3D Scene Reconstruction
    Éric Marchand
    François Chaumette
    International Journal of Computer Vision, 1999, 32 : 171 - 194
  • [33] Accurate 3D Reconstruction using a Turntable-based and Telecentric Vision
    Gonzalez-Barbosa, Jose-Joel
    Gomez-Loenzo, Roberto-Augusto
    Jimenez-Hernandez, Hugo
    Razo, Miguel
    Gonzalez-Barbosa, Ricardo
    AUTOMATIKA, 2015, 56 (04) : 508 - 521
  • [34] The path to accurate 3D data with handheld scanning solutions
    Patri, Gian-Philipp
    GIM INTERNATIONAL-THE WORLDWIDE MAGAZINE FOR GEOMATICS, 2024, 38 (05):
  • [35] On the data needed for the accurate reconstruction in 3D diffraction tomography
    Vouldis, AT
    Kechribaris, CN
    Maniatis, TA
    Nikita, KS
    Uzunoglu, NK
    2004 IEEE INTERNATIONAL WORKSHOP ON IMAGING SYSTEMS AND TECHNIQUES (IST), 2004, : 80 - 84
  • [36] Accurate attenuation modeling of rebinned 3D PET data
    Daube-Witherspoon, Margaret E.
    Matej, Samuel
    Vandenberghe, Stefaan
    Saffer, Janet R.
    Karp, Joel S.
    2005 IEEE NUCLEAR SCIENCE SYMPOSIUM CONFERENCE RECORD, VOLS 1-5, 2005, : 1910 - 1913
  • [37] Accurate loop calling for 3D genomic data with cLoops
    Cao, Yaqiang
    Chen, Zhaoxiong
    Chen, Xingwei
    Ai, Daosheng
    Chen, Guoyu
    McDermott, Joseph
    Huang, Yi
    Guo, Xiaoxiao
    Han, Jing-Dong J.
    BIOINFORMATICS, 2020, 36 (03) : 666 - 675
  • [38] 3D Vision in 3D Concrete Printing
    Sokolov, Dmitrii
    Mechtcherine, Viktor
    FOURTH RILEM INTERNATIONAL CONFERENCE ON CONCRETE AND DIGITAL FABRICATION, DC 2024, 2024, 53 : 182 - 189
  • [39] A Lens-Based Extension of Raycasting for Accurate Selection in Dense 3D Environments
    Plasson, Carole
    Cunin, Dominique
    Laurillau, Yann
    Nigay, Laurence
    HUMAN-COMPUTER INTERACTION, INTERACT 2021, PT IV, 2021, 12935 : 501 - 524
  • [40] Triangular mesh generation from 3D vision data
    Bradley, C
    Archibald, I
    Chan, VH
    THREE-DIMENSIONAL IMAGING AND LASER-BASED SYSTEMS FOR METROLOGY AND INSPECTION III, 1997, 3204 : 192 - 201