Dataset Viewer
The dataset viewer is not available for this subset.
Job has been terminated due to a temporary spike in resource usage and may be restarted later.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Scene dataset

Data is separated by subset. Currently available: train_1 and train_3. Future train_2 data will occupy a separate train_2/ directory. Scene names include the subset suffix, e.g. scene_000000_train1. The complete scene identity is (group, scene); every index and annotation records group.

The train_1 subset contains 17 generated tabletop scenes, 16 views per scene, 272 RGB-D observations, 16 unique object UIDs and 142 scene object instances.

Files

  • train_1/index.jsonl: one row per image, with scene/image indices, original reference view, object_uids, and relative file paths.
  • train_1/scene_*/realsense/: RGB, metric depth, instance labels, MAT object poses, camera calibration and per-image annotations/*.json.
  • train_1/scene_*/objects.json: object UID/label mapping, model-to-table poses and table geometry.
  • train_1/models/<uid>/: metre-scale GLB and original provenance metadata.
  • train_1/README_ZH.txt: detailed units, coordinate frames and reading instructions.

Object poses retain the GraspNet MAT convention: poses has shape 3x4xN, with model-to-camera rotation and translation in metres. cls_indexes maps each pose to its instance label. Per-image annotations provide object UIDs, these same object poses, and explicit camera/table homogeneous transforms. Table/camera JSON field names are documented dataset extensions.

RGB resolution is 1280x720. Depth PNGs are uint16 millimetres along the camera Z axis, with zero denoting invalid depth. Camera axes follow OpenCV. Scene/image indices start at zero; original scene numbers are preserved. UIDs identify the same object across scenes. Not every scene object is necessarily visible in every view; partial image cropping is allowed.

Status and limitations

For train_1, generation-stage checks completed. The separate batch verification was stopped at the user's request; this upload does not claim independent batch acceptance or training readiness.

Grasp stability, gripper collision and admissible opening-width supervision are not included. This is observation/geometry data, not a complete Graspformer training set or an official GraspNet dataset release.

Attribution and licenses

Object names, authors, source identifiers, URLs and original licenses are preserved in train_1/models/<uid>/metadata.json. Assets have individual terms; there is no replacement blanket license for the models. Consult those metadata files before reuse or redistribution.

train_3

The train_3 subset adds 38 scenes, 16 views per scene, 608 RGB-D observations and 32 unique object UIDs. Scene names end in _train3, for example train_3/scene_000000_train3/. The complete scene identity is (group, scene).

  • train_3/index.jsonl: image-level index; paths are relative to the repository root.
  • train_3/scene_*_train3/realsense/: RGB/depth/label PNGs, original GraspNet MAT, camera arrays and per-image annotations.
  • train_3/scene_*_train3/objects.json: UID/label mapping, model-to-table transforms and table geometry.
  • train_3/models/<uid>/: metre GLB and original author, license and source metadata.
  • train_3/README.txt: units, pose conventions and limitations.

Train_3 generation checks and independent batch geometry/physics verification passed. Copies were checked with SHA256; annotation poses match the original MAT. This does not constitute external asset review or grasp validation. Objects occur in 7-13 successful scenes; 11 objects have fewer than 10 appearances. Grasp stability, gripper collision and admissible opening-width supervision remain unavailable. This is not a complete grasp-training dataset. Model-specific licenses remain in each metadata.json; no blanket license replaces them. Logs, intermediate scene geometry, duplicate PLYs and environment assets are excluded from this delivery.

Portable scene-generation pipeline

The complete 15.05 GB runtime archive, English download/setup instructions, checksum, validation record, source notices, and AI deployment prompt are available in pipeline/. Start with the English guide and run one scene before generating a full subset. This is an unchanged portable snapshot from 2026-11-06; see its validation and license limitations.

Downloads last month
694