The 3DPW data files for this benchmark can be downloaded from their website alongside their images.
The in and out files contain the pose data for the indicated datasets. The first dimension of the data represents which video the data comes from, the second dimension determines the person, the third dimension indicates the time step, and the fourth dimension lists the coordinates of each of the joints.
The mask files contain the visibility data for the indicated datasets. This is only available for PoseTrack because 3DPW does not involve occluded joints. The shapes and formats of these files are the same as the pose data, except for the last dimension, which only includes one value for each joint (1 for visible and 0 for occluded).
The frame files list which video frames correspond to each input sequence of the pose data and can be used as inputs for your model.
This Python script contains the code for evaluating the VIM and VAM metrics on the predicted and ground truth pose sequences for a single person. These values should be averaged over all people in the dataset to obtain the scores shown on the leaderboard. Note that the 3DPW leaderboard is scaled by a value of 100.
Submit a .zip file containing files named 3dpw_predictions.json, posetrack_predictions.json, and posetrack_masks.json containing your model's predicted poses and visibility masks on the test set. These files should contain nested lists with the same format and dimensions as the supplied 3dpw_test_in.json, posetrack_test_in.json, and posetrack_test_masks_in.json files except that each sequence should contain 14 poses instead of 16. They should be in the root of your .zip file and not in any subdirectories. Please do not submit any additional files.
Note that information from the test output frames may not be used by your model in any way when generating predictions for this leaderboard.
You may submit to the leaderboard once every 24 hours.