pompomcork_YoloSegment-2D-to-3D-RebotARM

PompomCork YOLO-seg (tabletop)

YOLO instance-segmentation model for small tabletop objects (cork, lighter, pompom). It is the detector used by an RGB-D perception and grasping pipeline: 2D masks are lifted to 3D and turned into base-frame poses for robot picking.

• Task: instance segmentation
• Framework: Ultralytics YOLO
• Classes (ids): 0: cork, 1: lighter, 2: pompom
• Training image size: 448
• Training data: ddt1992/pompomcork_YoloSegment-2D-to-3D-RebotARM

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

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("image.jpg", task="segment", imgsz=448, conf=0.25)
for r in results:
    print(r.boxes.cls, r.masks is not None)

Intended use and limitations

• Trained on a small in-house dataset of three object classes on a wood table; it is not a general-purpose segmentation model and will not generalize to other objects or very different scenes/lighting.
• For symmetric objects (cork, pompom) orientation (yaw) is not meaningful and is not used downstream; the robot grasps them top-down.

Files

• best.pt: trained weights.

License

Released under the MIT License.