Robotic Garbage OBB Dataset (robotic-garbage-obb)

This dataset is designed for robotic parallel gripper grasping and top-down garbage classification using Oriented Bounding Boxes (OBB).

Images are captured in a real robotic bird's-eye view workspace (pure black background tabletop) using Intel RealSense cameras, labeled with oriented bounding boxes to provide both classification and exact planar yaw ($\theta$) orientation angles for robotic parallel grippers.

---

📸 Data Sources & Composition

The dataset contains 522 images curated specifically for robotic tabletop grasping:

• Real-world Robotic Captures:
  • Captured directly from the bird's-eye view camera of a robotic manipulation platform using Intel RealSense cameras against a matte black background.
  • Paired aligned depth maps (.npy, uint16 mm) and JET heatmaps are provided in demo/captures/.
• Stacked-Object Captures:
  • 51 RGB images of stacked / overlapping objects captured with an Intel RealSense D435 on a light wooden desk (oblique view, no depth), annotated with grasp-order relations.
  • 64 RGB-D captures of stacked objects (2026-10-26, RealSense D435IF on a tripod, top-down over a black cloth), annotated with grasp-order relations. Each has an aligned depth map in <split>/depth/ and intrinsics in <split>/metadata/.
• AI-Synthesized Augmentation (Google Nano Banana):
  • Part of the dataset images were generated using Google's Nano Banana AI model to enrich visual diversity, augment object orientation distributions, and balance class frequencies.
  • All AI-generated images underwent strict manual screening to ensure visual fidelity, realistic tabletop shadows, and consistent perspective alignment with physical robot workspace conditions.
• OBB Manual Annotation:
  • All images were verified and annotated with oriented bounding boxes (OBB) providing 4-corner coordinates and yaw grasping angles.

---

🏷️ Category Definitions & Gripper Physics Constraints

The dataset is categorized into 4 actionable robotic grasping classes:

Class ID	Class Name	Description	Robotic Grasping Strategy
0	plastic	Recyclable plastics (PET bottles, beverage cups, pudding cups, etc.)	Gripper aligns with the major/minor axis to enclose volume.
1	metal	Recyclable metal containers (aluminum cans, tin cans, coffee cans, etc.)	Aligns with cylindrical symmetry axis.
2	paper	Rigid structural paper containers (Tetra Pak foil cartons, fresh house milk cartons, instant noodle paper bowls, paper bento boxes)	Has structural rigidity and height for the gripper to close firmly.
3	general_waste	Non-recyclable or non-graspable waste (crumpled tissue paper, straws, plastic wrap, flat paper sheets / paper scraps)	Parallel grippers cannot grasp flat thin sheets resting directly on table surfaces; therefore, flat paper is classified as general_waste.

---

📁 Dataset Structure

robotic-garbage-obb/
├── dataset.yaml            # YOLO dataset configuration
├── train/                  # 381 images
│   ├── images/
│   ├── labels/
│   ├── grasp_order/        # grasp-order relations, one JSON per image
│   ├── depth/              # aligned depth for RGB-D captures: {stem}.npy (uint16 mm, 0 = invalid); only some images
│   └── metadata/           # {stem}.capture.json: camera serial, intrinsics, capture time
├── val/                    # 60 images
│   └── ...
├── test/                   # 81 images
│   └── ...
├── unlabeled/              # inbox for new captures (not used for training; empty when all are labeled)
│   └── ...                 # same subfolders; auto_label.py / label_tool.py (key T) work here, then files move to a split
├── demo/
│   ├── captures/           # RealSense pairs: {stem}_color.png, {stem}_depth.npy (uint16 mm), {stem}_depth_jet.png
│   └── captures_261026/    # same format; 10 samples from the 2026-10-26 stacked-object session
└── demo.gif                # Inference preview GIF

🔖 File Naming

Images are named by the first 8 hex characters of the SHA-256 of the file content (e.g. 004d33fb.jpg); the label, grasp-order, depth and metadata files share the same stem. Names never change, and the same image always gets the same name, so duplicates are detected on import.

Splits are fixed: existing images never move between train/val/test (so test stays unseen by earlier models); new images are assigned to a split when added.

📐 Annotation Format (YOLO-OBB)

Each label file contains 9 normalized space-separated values per object:

<class_index> <x1> <y1> <x2> <y2> <x3> <y3> <x4> <y4>

where $(x_1, y_1)$ through $(x_4, y_4)$ are the 4 clockwise corner coordinates normalized to $[0, 1]$.

Every visible object is labeled, including objects partly covered by others. Which object can be grasped first is stored separately in grasp_order/.

🔀 Grasp-Order Annotation (grasp_order/<stem>.json)

{"relations": [[2, 0], [2, 1]]}

• [up, down] = indices (line numbers, 0-based) in the label file: object up lies on / blocks object down, so up must be removed before down can be grasped safely. Drawn only for real physical constraints, never for merely adjacent objects; indirect ones (A on B on C ⇒ A before C) are implied.
• Images without any stacking have no JSON file.
• Objects with no incoming edge are graspable now (layer L1); removing them exposes L2, and so on.
• Currently 142 images with 493 relations. Annotated with label_tool.py (order mode, Tab).

---

🚀 Usage with Ultralytics YOLO11-OBB

from ultralytics import YOLO

# Train YOLO11-OBB model
model = YOLO("yolo11x-obb.pt")
model.train(data="dataset.yaml", epochs=300, imgsz=1024, batch=16)

---

📄 License

• Dataset annotations and images are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.