Hugging Face's logo

Datasets:
3dlg-hcvc
/
EgoFun3D

Languages:
English
Size:
n<1K
ArXiv:
Tags:
egocentric
3D
License:
Dataset card Files
xet
 Community

Acknowledge license to accept the repository

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By using data from this repo, users agree to the corresponding licensing of the original data: - Ego-Exo4D license. - FunGraph3D license (CC BY-NC 4.0).
To request access to EgoFun3D, you will need to provide your full name (please provide both your first and last name), the name of your advisor or the principal investigator (PI) of your lab (in the PI/Advisor) fields, and the school or company that you are affiliated with (the Affiliation field). After requesting access to this repo, you will be considered for access approval.

Log in or Sign Up to review the conditions and access this dataset content.

This dataset contains egocentric videos and paired interactive objects as described in EgoFun3D: Modeling Interactive Objects from Egocentric Videos using Function Templates. Some raw data is sourced from Ego-Exo4D, FunGraph3D, and some self-captured data. We thank the authors for open-sourcing their datasets.

File Structure

Our dataset follows this file structure

combined_dataset
     |--articulation_annotation
                  |--{object_id}.articulations.json
     |--cameras
           |--{video_name}.h5
     |--function_annotation
                |--{annotation_name}.json
     |--geometry
           |--{object_id}.ply or {object_id}.glb
     |--masks_h5
          |--{video_name}.h5
     |--object_transformations
              |--{object_id}.json
     |--part_annotation
              |--{object_id}.json
     |--videos
          |--{video_name}
                  |--undistorted.mkv or undistorted.mp4
                  |--cropped.mkv (when available)
     |--metadata.json
     |--merge_summary.json

The combined dataset contains 517 video sequences and 106 geometry assets. There are 517 undistorted videos and 226 cropped videos. Annotations and geometry can be shared by multiple sequences. File paths in metadata.json are relative to the dataset root and specify the exact filenames.

File Formats

Metadata

metadata.json contains a JSON list with one entry for each video sequence, defining all related annotations. Please see one example below.

{
    "video_name": "egoexo_82_2_0_indiana_cooking_10_2_keystep_492_Adjust-the-stove-heat_31",
    "video_mask": "masks_h5/egoexo_82_2_0_indiana_cooking_10_2_keystep_492_Adjust-the-stove-heat_31.h5",
    "camera": "cameras/egoexo_82_2_0_indiana_cooking_10_2_keystep_492_Adjust-the-stove-heat_31.h5",
    "mesh": "geometry/egoexo_new_82_2.ply",
    "geometry_type": "mesh",
    "object_transformation": "object_transformations/egoexo_new_82_2.json",
    "normalize_articulation_geometry": false,
    "articulation_in_canonical_frame": true,
    "object_mask": "part_annotation/egoexo_new_82_2.json",
    "articulation": null,
    "function": "function_annotation/egoexo_new_82_2_function.json",
    "function_id": "00",
    "base_type": "fixed",
    "initial_state": "close",
    "object_name": "stove",
    "object_id": "egoexo_new_82_2",
    "source": "egoexo4d",
    "undistorted_video": "videos/egoexo_82_2_0_indiana_cooking_10_2_keystep_492_Adjust-the-stove-heat_31/undistorted.mkv",
    "cropped_video": "videos/egoexo_82_2_0_indiana_cooking_10_2_keystep_492_Adjust-the-stove-heat_31/cropped.mkv"
}
  • video_name: the global unique identifier for video sequences.
  • source: the source of the video.
  • undistorted_video: the relative path to the undistorted MKV or MP4 video.
  • cropped_video: the relative path to the cropped MKV video, or null when no cropped video file exists.
  • video_mask: relative path to the HDF5 mask file of the video.
  • camera: relative path to the HDF5 file containing camera intrinsics and extrinsics.
  • mesh: relative path to the geometry file, including point clouds.
  • geometry_type: type of the geometry, can be "mesh" or "point cloud".
  • object_transformation: relative path to the transformation from canonical geometry coordinates to the source world/camera frame.
  • object_mask: relative path to the geometry part annotation file.
  • articulation: relative path to the articulation annotation file, or null when no annotation is available.
  • function: relative path to the part functionality annotation file.
  • function_id: the string identifier for the part functionality for the corresponding geometry. Preserve leading zeros when present.
  • object_name: name of the object.
  • object_id: the global unique identifier for the object.
  • base_type: base type of the object.
  • initial_state: annotated initial state of the object to compute the initial joint value.
  • articulation_in_canonical_frame: whether articulation parameters are expressed in the canonical geometry frame.
  • normalize_articulation_geometry: whether the loader adjusts part geometry to joint minima. This is false for the old dataset to preserve the observed articulation pose; its geometry is still recentered and its articulation annotations are in the canonical frame.

Articulation

Each object with articulation annotations has an annotation file, containing annotated parameters for every articulated joint of the object. Files contain either a joint list or a JSON object with an articulations list. One articulated joint annotation example is shown below.

[
    {
        "pid": 1,
        "type": "revolute",
        "base": [
            0
        ],
        "axis": [
            0.7618739307067427,
            -0.6477253381715566,
            -1.4382391681423592e-16
        ],
        "origin": [
            0.025152706356387444,
            -0.03601834938695314,
            -0.06390552222728724
        ],
        "rangeMin": 0,
        "rangeMax": 1.5707963267948966
    }
]
  • pid: part id of the child link on the mesh/point cloud annotation, corresponding to the part annotation file.
  • type: joint type.
  • base: parent link id.
  • axis: joint axis in the canonical geometry frame.
  • origin: joint position for revolute joint in the canonical geometry frame.
  • rangeMin: minimum joint value.
  • rangeMax: maximum joint value.

Revolute ranges are in radians; prismatic ranges use the geometry's distance unit. Additional fields such as ref and motionStates are retained when available.

Camera

Each cameras/{video_name}.h5 file stores camera intrinsics and extrinsics together:

{video_name}.h5
     |--T_world_camera         (N, 4, 4)
     |--K                      (3, 3)
     |--K_cropped              (3, 3), optional
     |--cropped_top_left       (2,), optional
     |--cropped_bottom_right   (2,), optional
     |--original_frame_size    (2,), optional
  • T_world_camera: camera-to-world matrices in OpenCV convention. N is the number of frames.
  • K: intrinsic matrix for undistorted frames.
  • K_cropped: intrinsic matrix for cropped frames, preserved when crop calibration is available.
  • cropped_top_left, cropped_bottom_right: top-left and exclusive bottom-right crop coordinates (u, v) on the undistorted frames.
  • original_frame_size: (H, W) for the undistorted frames.

Some captures may also contain T_world_device, T_device_camera, capture_timestamp_ns, frame_index, and pose_valid.

Function

Each object has a function annotation file, containing a dictionary of annotated function templates keyed by string function id. One function template annotation example is shown

{
    "00": {
        "physics": "geometry",
        "func": "linear"
    }
}
  • physics: physical effect.
  • func: mapping.
  • 00: function instance id for the specific scene/object, corresponding to function_id in the metadata.

Geometry

It is a reconstruction of scene/object. The ply file is either a point cloud or a mesh; glb files contain meshes.

All meshes are recentered at the origin. Use the corresponding object transformation to place geometry in its source coordinate system.

Object Transformation

Each object_transformations/{object_id}.json file contains a canonical_to_world 4-by-4 homogeneous matrix, stored as a row-major nested list, together with transformation and coordinate-system information.

For homogeneous column vectors:

p_world = canonical_to_world @ p_canonical
p_camera = inverse(T_world_camera[frame]) @ p_world

The coordinate_system and matrix_convention fields describe the target frame. Use the geometry path in metadata.json; embedded provenance paths can refer to the pre-merge layout.

Mask

Stored in a HDF5 file with the following structure. Every annotation group contains one dataset named mask and has id and name attributes.

{video_name}.h5
     |--left hand (when available)
           |--attributes: id=1, name="left hand"
           |--mask (N, H, W)
     |--right hand (when available)
           |--attributes: id=2, name="right hand"
           |--mask (N, H, W)
     |--receptor
           |--attributes: id=3, name={part_name}
           |--mask (N, H, W)
     |--effector
           |--attributes: id=4, name={part_name}
           |--mask (N, H, W)
     |--object (when available)
           |--attributes: id=5, name={object_name}
           |--mask (N, H, W)
     |--additional part1 (when available)
           |--attributes: id=6, name={additional_part_name}
           |--mask (N, H, W)
     |--additional part2, ... (when available)
           |--attributes: id={mask_id}, name={additional_part_name}
           |--mask (N, H, W)
  • name: original name of the annotated part/object, stored as a group attribute. Additional annotations retain names such as "gas lighter" or "drawer interior" here.
  • id: mask id, stored as a group attribute. IDs 1 and 2 identify hands, 3 the receptor, 4 the effector, and 5 the whole object. Existing additional-part IDs are retained; missing IDs are assigned unused values starting at 6.
  • mask: frame masks of shape (N, H, W) in undistorted video coordinates; nonzero values indicate foreground. This dataset name is used consistently for every group.
  • additional part1, additional part2, etc.: numbered within each file. Original named groups are ordered alphabetically when assigning these names.

Part

Part annotation files contain a list of records. Select records whose function_instance_id matches the metadata's function_id. The example below shows abbreviated vertex-index lists.

[
    {
        "function_instance_id": "01",
        "scene_id": "6",
        "description": "control the water flow",
        "receptor": {
            "label": "faucet handle",
            "indices": [641, 642, 643, 644, 645],
            "pid": 1
        },
        "effector": {
            "label": "faucet spout",
            "indices": [2868, 2985, 2986, 2987, 3103],
            "pid": 2
        }
    }
]
  • function_instance_id: function instance id for the specific scene/object.
  • scene_id: associated scene/object id.
  • description: natural language description of the part functionality.
  • label: part name.
  • indices: vertex indices for the mesh or point indices for the point cloud for the part. GLB vertex ordering follows the benchmark loader's scene flattening.
  • pid: id for this part.

Video

  • undistorted.mkv or undistorted.mp4: undistorted video.
  • cropped.mkv: cropped video from the undistorted video to remove black areas, when available.

Citation

If you find our dataset to be useful, please cite our paper

@article{peng2026egofun3d,
  title={{EgoFun3D: Modeling Interactive Objects from Egocentric Videos using Function Templates}},
  author={Peng, Weikun and Iliash, Denys and Savva, Manolis},
  journal={arXiv preprint arXiv:2604.11038},
  year={2026}
}
Downloads last month
232
Total file size:
71.7 GB

Paper for 3dlg-hcvc/EgoFun3D