Real test dataset with manually refined ground truth labels for human instance and body part segmentation

Based on the data from the EgoBody dataset, we created a point-cloud dataset with manually annotated human instance labels and refined body part labels. Our test dataset is currently hosted on the original EgoBody download page, and can be downloaded upon accepting the license agreement. On the EgoBody download page, please download the human3d_egobody_test_set_release.zip file to access this dataset, which consists of 304 point clouds along with their ground truth per-point annotations for human instances and body parts. The overall zipped size of the annotated test scenes is less than 1G.

Real training/validation dataset with pseudo-GT labels for human instance and body part segmentation

Based on the data from the EgoBody dataset, we created a point-cloud dataset with pseudo GT per-point labels for human instance and body part segmentation tasks. Our dataset is currently hosted on the original EgoBody download page, and can be downloaded upon accepting the license agreement. On the EgoBody download page, please download the human3d_egobody_pcd_data.zip file to access this dataset, which is approximately ~62G, and includes around 20k scenes (point clouds) for training and validation.

Synthetic Dataset

For instructions regarding the synthetic training data generation, please see https://github.com/human-3d/SyntheticHumanDataset.

Loading and Visualizing Data

We provide an example dataloader for loading and visualizing the point clouds in our dataset. The example dataloader can be found at https://github.com/human-3d/SyntheticHumanDataset/blob/main/example_dataloader.py. Please see the README for further details about how to set up the repo, and how to use the dataloader.