Datasets:

gserifi
/

SynthesiaSet

dataset_info:
  features:
    - name: image
      dtype: image
    - name: kpts
      struct:
        - name: depth
          list: float64
        - name: keypoints2d
          list:
            list: float64
        - name: keypoints3d
          list:
            list: float64
  splits:
    - name: train
      num_bytes: 8215932300
      num_examples: 100000
  download_size: 8267044940
  dataset_size: 8215932300
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: apache-2.0
task_categories:
  - object-detection
pretty_name: SynthesiaSet
size_categories:
  - 100K<n<1M

SynthesiaSet

Matan Davidi Cyril Moser Gent Serifi Nicola Studer

ETH Zürich, Switzerland

[Project on GitHub]

Synthetic dataset containing 100K Piano images with 2D+3D keypoint annotations (the 4 corners), rendered with Mitsuba 3.

Created as part of our course project for Mixed Reality at ETH Zürich.

image: RGB, 224x304
kpts: dictionary containing:
- keypoints3d: (4, 3) 3D positions for all 4 keypoints (in clockwise order, starting from the top-left)
- keypoints2d: (4, 2) 2D keypoint projections
- depth: (4,) Distance to the camera

References

@software{jakob2022mitsuba3,
    title = {Mitsuba 3 renderer},
    author = {Wenzel Jakob and Sébastien Speierer and Nicolas Roussel and Merlin Nimier-David and Delio Vicini and Tizian Zeltner and Baptiste Nicolet and Miguel Crespo and Vincent Leroy and Ziyi Zhang},
    note = {https://mitsuba-renderer.org},
    version = {3.0.1},
    year = 2022,
}