Dataset preparation¶
Overview
Supported datasets
Download converted meta-data
Convert a dataset manually
Validate converted meta-data by visualization
Overview¶
Our data pipeline converts original dataset to our unified meta-data, with data converters controlled by configs.
Supported datasets¶
| Dataset name | Dataset page | Download |
|---|---|---|
| Campus | Home page | CampusSeq1.tar.bz2 |
| Shelf | Home page | Shelf.tar.bz2 |
| CMU Panoptic | Home page | By official script |
| FourDAG | Home page | dataset (password in home page) |
Download converted meta-data¶
Considering that it takes long to run a converter if perception2d is checked, we have done it for you. Our perception 2D is generated by mmtrack and mmpose, defined in coco_wholebody by default. You can download compressed zip file for converted meta-data below.
For where to put the downloaded meta-data, check xrmocap dataset structure for details.
| Dataset name | meta name | Download link | Notes |
|---|---|---|---|
| Campus | testset | download | Paf data is included. |
| Campus | testset_fasterrcnn | download | Bbox 2D is generated by mmdet Faster R-CNN. |
| Campus | testset_mvpose2d | download | Perception 2D is generated by MVPose, defined in coco convention. |
| Campus | trainset | download | |
| Campus | trainset_pesudo_gt | download | Ground-truth keypoints3d is generated by MvP, defined in campus convention. |
| Shelf | testset | download | Paf data is included. |
| Shelf | testset_fasterrcnn | download | Bbox 2D is generated by mmdet Faster R-CNN. |
| Shelf | testset_mvpose2d | download | Perception 2D is generated by MVPose, defined in coco. There's only data for the first three people in ground truth keypoints3d . |
| Shelf | trainset | download | |
| Shelf | trainset_pesudo_gt | download | Ground-truth keypoints3d is generated by MvP, defined in campus convention. |
| CMU Panoptic | 160906_band4 | download | Only five views are selected: 03, 06, 12, 13, 23 |
| CMU Panoptic | 160422_haggling1 | download | Only five views are selected: 03, 06, 12, 13, 23 |
| CMU Panoptic | 160906_ian5 | download | Only five views are selected: 03, 06, 12, 13, 23 |
| CMU Panoptic | 160906_pizza1 | download | Only five views are selected: 03, 06, 12, 13, 23 |
| FourDAG | seq2 | download | Paf data is included. |
| FourDAG | seq4 | download | Paf data is included. |
For CMU panoptic meta-data, frames extracted from videos have been removed before uploading. One has to convert panoptic data locally with bbox_detector = None and kps2d_estimator = None first, and then copy download data into the converted meta-data directory.
Convert a dataset manually¶
Use our prepare_dataset tool to convert a dataset. See the tool tutorial for details.
Validate converted meta-data by visualization¶
Use our visualize_dataset tool to visualize meta-data. See the tool tutorial for details.
For CMU panoptic meta-data, the multi-view all-in-one video is too large to load, please set vis_aio_video = False to avoid OOM.