ddt1992/pompomcork_YoloSegment-2D-to-3D-RebotARM

PompomCork RGB-D Tabletop Segmentation (2D-to-3D)

Instance-segmentation dataset of small tabletop objects (cork, pompom, lighter) captured with an Intel RealSense camera. It is the data behind an RGB-D perception and grasping pipeline: 2D YOLO instance segmentation, lifting masks to 3D, and base-frame pose estimation for robot picking.

 

Code, training and the full pipeline (ROS 2 / MoveIt): https://github.com/danieldoradotalaveron-rb/YoloSegment-2D-to-3D-RebotARM_Pick_and_Place

Model (trained weights): https://huggingface.co/ddt1992/pompomcork_YoloSegment-2D-to-3D-RebotARM

What this dataset contains

• Raw images (raw/): 169 pristine RGB captures, no labels.
• Labeled images (labeled/): 169 RGB images with instance masks in LabelMe polygon format (.png/.jpg image + matching .json), in two passes: manual_labelme/ (first-pass hand labeling) and reviewed_yolo_labelme/ (YOLO-assisted, reviewed).
• Synthetic, train-only (synthetic_point/, optional): 114 composited views generated by lifting labeled objects to 3D and re-rendering them.
• RGB-D source (rgbd/, rgbd_backgrounds/, optional): 23 RGB-D object captures and 23 empty-scene backgrounds, enough to regenerate the synthetic set.
• Ready-to-train YOLO-seg export (yolo/): the converted dataset (270 train / 68 val images, polygon-mask .txt labels and a data.yaml). This is a build artifact regenerated from the labels above by just convert on the GitHub side; it is included here for convenience.
• Classes: cork, lighter, pompom (class ids 0/1/2 in the YOLO export).

Structure

raw/                    # 169 raw RGB images (no labels)
labeled/                # reviewed LabelMe pairs (image + .json)
  manual_labelme/       #   first-pass hand labeling
  reviewed_yolo_labelme/#   YOLO-assisted, reviewed
synthetic_point/        # (optional) 114 synthetic composited pairs, train-only
rgbd/                   # (optional) 23 RGB-D object captures
rgbd_backgrounds/       # (optional) 23 empty-scene RGB-D backgrounds
yolo/                   # ready-to-train YOLO-seg export (regenerable via `just convert`)
  data.yaml             #   class names: cork, lighter, pompom
  images/{train,val}/   #   270 train / 68 val images
  labels/{train,val}/   #   polygon-mask .txt labels

Intended use

• Training and evaluating instance-segmentation models for small tabletop objects.
• Feeding an RGB-D 2D-to-3D lifting and pick-and-place pipeline.

The validation split should stay 100% real. Any synthetic augmentation is generated and validated on the GitHub side.

How it was produced

Captured with an Intel RealSense camera, labeled in LabelMe (first pass by hand, later assisted by a trained YOLO model and reviewed). Full capture, labeling and training workflow: see the GitHub repository linked above.

License

Released under the MIT License.