dccc2025/ovok_experimental_datasets

OVOK DREAM-V and RoboKeyGen-V

Packed paper-table sets for OVOK. Each clip is one folder: RGB video plus one gt.json.

These are not new captures. They are processed motion clips from DREAM and RoboKeyGen. Cite the original papers if you use this data.

Layout

dream_v/<clip_id>/video.mp4
dream_v/<clip_id>/gt.json
robokeygen_v/<clip_id>/video.mp4
robokeygen_v/<clip_id>/gt.json

Set	Clips	Frames	Resolution
dream_v	60	37379	640×480
robokeygen_v	24	9283	640×360

gt.json uses the OVOK scorer schema. video is <clip_id>.mp4. frame_index starts at 0 and is consecutive in the same order as the mp4. Each frame has keypoints_2d, keypoints_3d, and joint_angles_deg for panda_link0/2/3/4/6/7/hand and panda_joint1–panda_joint7.

RoboKeyGen 2D is location_wrt_cam projected with the official scene K.

Sources and citations

DREAM (dream_v/)

Lee, T. E., Tremblay, J., To, T., Cheng, J., Mosier, T., Kroemer, O., Fox, D., and Birchfield, S.
Camera-to-Robot Pose Estimation from a Single Image. ICRA 2020.
https://arxiv.org/abs/1911.09231
Code/data: https://github.com/NVlabs/DREAM

Processed from the official real Panda splits (panda-3cam_azure, panda-3cam_kinect360, panda-3cam_realsense, panda-orb). Capture rate in DREAM is 30 fps, 640×480.

RoboKeyGen (robokeygen_v/)

Tian, Y., Zhang, J., Huang, G., Wang, B., Wang, P., Pang, J., and Dong, H.
RoboKeyGen: Robot Pose and Joint Angles Estimation via Diffusion-based 3D Keypoint Generation. ICRA 2024.
https://arxiv.org/abs/2403.18259
Project: https://nimolty.github.io/Robokeygen/

Processed from Real_Test_0613.tar.gz (RealSense-Franka and AzureKinect-Franka). Official test scenes only plus extra scenes 7_kinect_left_0 and 11_D415_right_0. RGB is 640×360. Clips are encoded at 30 fps.