THEODORE+

News

2020-04-27 - THEODORE+ Datensatz mit Bildern, Begrenzungsrahmen, Masken, Körperkeypoints und Aktionen veröffentlicht.
2023-04-05 - Paper von CVPR OmniCV 2023 workshop akzeptiert.

Abstract

The implementation of THEODORE+ closely follows THEODORE, utilizing six indoor settings and domain randomization to create a diverse dataset. The object textures and the human model parameters (height and weight) are randomized, and at the same time the camera position changes constantly to ensure varied perspectives. During the simulation, each character has eight pre-defined animations that belong to four categories: sitting, lying, falling and walking. An animation is selected from this animation set and executed depending on the character’s action. The corresponding action of a virtual character is also part of the exported data and is usable for activity recognition. Beside bounding boxes and segmentation masks, our dataset features full body 2D and 3D pose information and action information of the human model. The implementation based on the Unity engine is extended to map the human body skeleton points from the engine’s internal format to the COCO format with 13 keypoints. No data is available for the simulated characters for the keypoints left eye, right eye, left ear and right ear. Therefore, these points receive the coordinates [0,0] during the export. THEODORE+ consists of 50,000 images with a resolution of 2048×2048 pixel and ∼160,000 character instances. Each file is saved in PNG format to exclude compression artifacts in the exported images. The 2D keypoints are converted to pixel coordinate space during export, while the 3D keypoints remain unchanged.

Paper

BibTeX

Wenn Sie den Datensatz in Ihrer Arbeit verwenden, bitte folgende Referenz nicht vergessen:

@InProceedings{Yu_2023_CVPR,
    author    = {Yu, Jingrui and Scheck, Tobias and Seidel, Roman and Adya, Yukti and Nandi, Dipankar and Hirtz, Gangolf},
    title     = {Human Pose Estimation in Monocular Omnidirectional Top-View Images},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2023},
    pages     = {6410-6419}
}

Lizenz

Der Datensatz ist lizenziert unter einer Creative Commons Namensnennung 4.0 International Lizenz.