We have annotated several videos to create a new ground contact dataset, including training, validation, and test splits. You can download them using Google Drive or VT server.
contact_dataset: test split
- labels: ground contact labels (fully annotated)
- openpose: original OpenPose results
- openpose_flow: parsed OpenPose results
- videos: input video
Human3.6M: including training and validation videos
- noisy_ground_2: ground contact labels (partially annotated, "-1" means not annotated)
- openpose: parsed OpenPose results
MADS: including training videos
- noisy_ground_2_mads: ground contact labels (partially annotated, "-1" means not annotated)
- openpose: parsed OpenPose results
We use four numbers to represent the ground contact label for Left Toe, Left Heel, Right Toe, Right Heel
. 0 means not in contact, 1 means in contact, -1 means not annotated.
Some of the videos from Human3.6M or MADS.
NOTE: Due to data license of Human3.6M, we will not provide any visual data for the videos coming from this dataset. Instead, we only provide the pre-computed OpenPose resutls and ground contact labels.