Contains the synthetic dataset developed for our CVPR2021 paper:

🚀 A Monocular Pose Estimation Case Study: The Hayabusa2 Minerva-II2 Deployment. 🚀

Images shared under CC0.

Synthetic Minerva-II2 images currently available for download at Price's Google Drive

The real Minerva-II2 images are copyright to the Hayabusa 2 Optical Navigation Camera (ONC) system team: JAXA, Chiba Institute of Technology, University of Tokyo, Kochi University, Rikkyo University, Nagoya University, University of Aizu and AIST. For access to the real Minerva-II2 deployment images permission will need to be obtained from JAXA.

For any questions or difficulties regarding accessing the dataset, please feel free to contact the authors. [email protected]

Supplementary Video: https://youtu.be/87R-tGOvD4U

Three annotation files are shared. Their content format is identical.

1. SetA - A set of ~10,000 images with a symmetric Minerva-II2 target. The set was designed to replicate real Minerva-II2 deployment images.
2. SetB - A set of ~10,000 images with a nonsymmetric Minerva-II2 target. The set was designed to work with the Tumble dataset.
3. Tumble - A set of 300 images with a nonsymmetric Minerva-II2 target. The set was designed to challenge a pose estimation pipeline to recover and estimate tumbling motion.

The annotation files are formatted with the following members:

• image: image name
• joints: pixel locations of vertices [pixels, pixels]
• pose: camera frame pose of the Minerva-II2 target [q0,q1,q2,q3,m,m,m]
• Example entry:

'{
  "image":"image_1.png",
  "joints":[[1,1],[2,2],[3,3],[4,4],[5,5],[6,6],[7,7],[8,8],[9,9],[10,10],[11,11],[12,12],[13,13],[14,14],[15,15],[16,16]],
  "pose":[1,0,0,0,0,0,0.5]
}'

The "joints" member contains 16 [x,y] pairs indicating the location of a vertice in the image. The "pose" member is a 1x7 vector containing the camera frame pose of the Minerva-II2 (quaternion|cartesian vector).

If desired, one can generate their own "joints" annotations for training based on the pose. We include the joints information here for convenience. Please note, the datasets have a slight distinction:

• The SetA "pose" [cartesian] vector points to the [0.875, -0.04, 4.76]_{MinervaII2_bodyFrame}
• The SetB "pose" [cartesian] vector points to the [0, 0, 0]_{MinervaII2_bodyFrame}
• The Tumble "pose" [cartesian] vector points to the [0, 0, 0]_{MinervaII2_bodyFrame}

The geometry of the Minerva-II2 vertices are shared in MinervaII2_Vertices.txt. The values are in [mm].

The renderings were designed to match the Hayabusa2 ONC-W2 camera parameters.

• resolution: 1024x1024
• sensor size: 13um x 13um
• focal length: 10.38 mm
• field of view: 68.89 deg

The dataset development is described in our paper.

@inproceedings{Price21,
  title={A Monocular Pose Estimation Case Study: The Hayabusa2 Minerva-II2 Deployment},
  author={Price, Andrew and Yoshida, Kazuya},
  booktitle={CVPR},
  year={2021}
}