3D human pose estimation with siamese equivariant embedding

Márton Véges, Viktor Varga, A. Lőrincz

Research output: Article

Abstract

In monocular 3D human pose estimation a common setup is to first detect 2D positions and then lift the detection into 3D coordinates. Many algorithms suffer from overfitting to camera positions in the training set. We propose a siamese architecture that learns a rotation equivariant hidden representation to reduce the need for data augmentation. Our method is evaluated on multiple databases with different base networks and shows a consistent improvement of error metrics. It achieves state-of-the-art cross-camera error rate among algorithms that use estimated 2D joint coordinates only.

Original languageEnglish
Pages (from-to)194-201
Number of pages8
JournalNeurocomputing
Volume339
DOIs
Publication statusPublished - ápr. 28 2019

Fingerprint

Cameras
Joints
Databases

Keywords

    ASJC Scopus subject areas

    • Computer Science Applications
    • Cognitive Neuroscience
    • Artificial Intelligence

    Cite this

    3D human pose estimation with siamese equivariant embedding. / Véges, Márton; Varga, Viktor; Lőrincz, A.

    In: Neurocomputing, Vol. 339, 28.04.2019, p. 194-201.

    Research output: Article

    Véges, Márton ; Varga, Viktor ; Lőrincz, A. / 3D human pose estimation with siamese equivariant embedding. In: Neurocomputing. 2019 ; Vol. 339. pp. 194-201.
    @article{cb408c213e524ed894164ed88a9ec8c4,
    title = "3D human pose estimation with siamese equivariant embedding",
    abstract = "In monocular 3D human pose estimation a common setup is to first detect 2D positions and then lift the detection into 3D coordinates. Many algorithms suffer from overfitting to camera positions in the training set. We propose a siamese architecture that learns a rotation equivariant hidden representation to reduce the need for data augmentation. Our method is evaluated on multiple databases with different base networks and shows a consistent improvement of error metrics. It achieves state-of-the-art cross-camera error rate among algorithms that use estimated 2D joint coordinates only.",
    keywords = "3D pose estimation, Equivariant embedding, Siamese network",
    author = "M{\'a}rton V{\'e}ges and Viktor Varga and A. Lőrincz",
    year = "2019",
    month = "4",
    day = "28",
    doi = "10.1016/j.neucom.2019.02.029",
    language = "English",
    volume = "339",
    pages = "194--201",
    journal = "Neurocomputing",
    issn = "0925-2312",
    publisher = "Elsevier",

    }

    TY - JOUR

    T1 - 3D human pose estimation with siamese equivariant embedding

    AU - Véges, Márton

    AU - Varga, Viktor

    AU - Lőrincz, A.

    PY - 2019/4/28

    Y1 - 2019/4/28

    N2 - In monocular 3D human pose estimation a common setup is to first detect 2D positions and then lift the detection into 3D coordinates. Many algorithms suffer from overfitting to camera positions in the training set. We propose a siamese architecture that learns a rotation equivariant hidden representation to reduce the need for data augmentation. Our method is evaluated on multiple databases with different base networks and shows a consistent improvement of error metrics. It achieves state-of-the-art cross-camera error rate among algorithms that use estimated 2D joint coordinates only.

    AB - In monocular 3D human pose estimation a common setup is to first detect 2D positions and then lift the detection into 3D coordinates. Many algorithms suffer from overfitting to camera positions in the training set. We propose a siamese architecture that learns a rotation equivariant hidden representation to reduce the need for data augmentation. Our method is evaluated on multiple databases with different base networks and shows a consistent improvement of error metrics. It achieves state-of-the-art cross-camera error rate among algorithms that use estimated 2D joint coordinates only.

    KW - 3D pose estimation

    KW - Equivariant embedding

    KW - Siamese network

    UR - http://www.scopus.com/inward/record.url?scp=85062100251&partnerID=8YFLogxK

    UR - http://www.scopus.com/inward/citedby.url?scp=85062100251&partnerID=8YFLogxK

    U2 - 10.1016/j.neucom.2019.02.029

    DO - 10.1016/j.neucom.2019.02.029

    M3 - Article

    VL - 339

    SP - 194

    EP - 201

    JO - Neurocomputing

    JF - Neurocomputing

    SN - 0925-2312

    ER -