HOH-Grasps Dataset

The dataset contains three main zip files: object_grasp_labels.zip, which stores all object grasp labels, and scenes_group1.zip and scenes_group2.zip, which store all scene data. All units are in meters, and all angles are in radians.

object_grasp_labels

The object_grasp_labels.zip when unzipped yields the folder object_grasp_labels under which, there are three subfolders:

1. graspnet: GraspNet-format [1] grasp labels stored in .npz files, named by object IDs from the HOH dataset [2]. Each file contains the following keys:
   • points: Sampled grasp point positions. Shape: N × 3, where N is the number of grasp points.
   • offsets: Offsets for each grasp. Shape: N × 300 × 12 × 4 × 3, where dimensions correspond to 300 viewpoints, 12 in-plane rotation angles, and 4 grasp depths. The last dimension stores [angle, depth, width] for each grasp.
   • collision: Grasp collision labels (True if grasp is invalid due to collision, otherwise False). Shape: N × 300 × 12 × 4.
   • scores: Grasp friction coefficients μ (lower scores indicate higher grasp quality). Valid grasps range [0.1, 1]; invalid grasps are set to -1. Shape: N × 300 × 12 × 4.
   • tolerance: Tolerance scores. Shape: N × 300 × 12 × 4.
2. plain_6d_valid: Plain 6D valid grasp labels stored in .npz files, named by object IDs from the HOH dataset. Labels are derived from GraspNet grasp labels with invalid grasps removed. Each file contains the following keys:
   • translations: Grasp translation vectors. Shape: N × 3.
   • rotation_matrices: Grasp rotation matrices. Shape: N × 3 × 3.
   • widths: Gripper widths. Shape: N.
   • depths: Gripper depths. Shape: N.
   • scores: Grasp quality scores (S = 1.1 - μ; higher is better). Range: [0.1, 1]. Shape: N.
   • tolerance: Tolerance scores. Shape: N.
3. plain_6d_good: A subset of plain_6d_valid containing only good grasps with S >= 0.6.

And four additional files:

1. object_ids.json: Contains all 136 object IDs from the HOH dataset.
2. object_metadata.csv: Provides information in tabular format for all 136 objects. Row headers correspond to object ID, a short text description, mass (g), vertex count, face count, function category, aspect ratio category, and the number of total grasps, valid grasps, and good grasps.
3. scene_ids.json: Contains three keys:
   • all: List of all 2,720 scene IDs.
   • train: List of 1,700 training scene IDs.
   • test: List of 1,020 testing scene IDs.
4. scene_metadata.csv: Provides information in tabular format for all 2,720 scenes. Row headers correspond to scene ID, giver ID, receiver ID, object ID, the number of plain 6D valid grasps at the O and T frames, and the number of plain 6D good grasps at the O and T frames.

scenes

The scenes_group1.zip files contains one group 1,360 scene subfolders, and scenes_group1.zip files contains a second group of 1,360 scene subfolders corresponding to the scenes in the HOH dataset. It is recommended to unzip the two files, and put all the resultant 2,720 subfolders into a single folder called scenes at the same level as the object_grasp_labels and the four json/csv files. Scenes are rotated from the HOH dataset so that the giver is positioned on the –Y side and the receiver on the +Y side. The handover direction is towards +Y. Each scene subfolder is named [scene ID], formatted as [giver ID]_[receiver ID]-S[starting interaction number]_[interaction index]_[object ID]. For example, 02781-96731-S1_0_308 represents the first interaction (0) from the recording starting with interaction number 1 (S1) between giver 02781 and receiver 96731, using object 308. Data is provided for two key frames:

• Original frame (O Frame): Moment immediately before human grasping (corresponding to the Gpre or O frame in HOH).
• Transfer Frame (T Frame): Moment of transferring the object during handover, when both giver's and receiver's hands contact the object (corresponding to T frame in HOH).

Each scene subfolder contains:

1. [scene ID]_scene_o.ply: Combined object-tabletop point cloud at the O frame.
2. [scene ID]_scene_t.ply: Combined object-tabletop point cloud at the T frame.
3. [scene ID]_tabletop.ply: Tabletop point cloud.
4. [scene ID]_hand_giver_o.ply: Giver’s hand point cloud from the G frame in HOH, aligned to the O frame.
5. [scene ID]_hand_giver_t.ply: Giver’s hand point cloud at the T frame.
6. [scene ID]_hand_receiver_o.ply: Receiver’s hand point cloud from the T frame in HOH, aligned to the O frame.
7. [scene ID]_hand_receiver_t.ply: Receiver’s hand point cloud at the T frame.
8. [scene ID]_obj_o.ply: Object-only point cloud at the O frame.
9. [scene ID]_obj_t.ply: Object-only point cloud at the T frame.
10. [scene ID]_grasps_valid_o.npz: plain_6d_valid grasp labels for this scene at the O frame (tabletop collisions removed).
11. [scene ID]_grasps_valid_t.npz: plain_6d_valid grasp labels for this scene at the T frame (tabletop collisions removed).
12. [scene ID]_grasps_good_o.npz: plain_6d_good grasp labels for this scene at the O frame (tabletop collisions removed).
13. [scene ID]_grasps_good_t.npz: plain_6d_good grasp labels for this scene at the T frame (tabletop collisions removed).
14. [scene ID]_objectness.npz: Objectness mask differentiating object and tabletop in the scene point clouds. Keys:
    • o: Objectness mask for [scene ID]_scene_o.ply.
    • t: Objectness mask for [scene ID]_scene_t.ply.
15. [scene ID]_scene_collision.npz: GraspNet-format collision labels with tabletop collisions marked as True. Keys:
    • o: Collision labels at O frame. Shape: N × 300 × 12 × 4.
    • t: Collision labels at T frame. Shape: N × 300 × 12 × 4.
16. [scene ID]_centroids.json: Centroids, including the following keys:
    • obj_o: Object centroid at the O frame, obtained by aligning the object's mesh centroid to the object pose at the O frame.
    • obj_t: Object centroid at the T frame, obtained by aligning the object's mesh centroid to the object pose at the T frame.
    • hand_giver_o: Centroid of the giver's hand point cloud at the O frame.
    • hand_giver_t: Centroid of the giver's hand point cloud at the T frame.
    • hand_receiver_o: Centroid of the receiver's hand point cloud at the O frame.
    • hand_receiver_t: Centroid of the receiver's hand point cloud at the T frame.
17. [scene ID]_transforms.json: Transformations, including the following keys:
    • o_to_t: Transformation from the O frame to the T frame.
    • t_to_o: Transformation from the T frame to the O frame.
    • obj_to_o: Transformation from the object mesh coordinate frame to the O frame.
    • obj_to_t: Transformation from the object mesh coordinate frame to the T frame.

References

[1] H.-S. Fang et al., "Graspnet-1billion: A large-scale benchmark for general object grasping," CVPR, 2020, pp. 11,444–11,453.

[2] N. Wiederhold et al., "HOH: Markerless multimodal human-object-human handover dataset with large object count," NeurIPS, vol. 36, 2023, pp. 68,736–68,748.