README.md

metadata

license: etalab-2.0
rai:
  dataLimitations: >-
    TiBuDB contains images from different gas-liquid mass transfer experiment
    but not the attached explotary parameter set (e.g. pressure or gas/liquid
    flow rate). Our strategy is annotating based on smaller patches then infer
    on complete image for reducing annotating effort. Consequently, bubbles
    located at the image edges may appear  as standard horizontal bounding boxes
    rather than oriented bounding boxes in case of oriented detection. Because
    the rotation angles are difficult to determine when the bubble is truncated
    by the frame when patching. Also, some regions with low contrast or strong
    overlap may lead to less accurate annotations.
  dataBiases: >-
    TiBuDB focus on tiny bubble detection, because the larger scale bubble now
    is not integrate in our set it may lead to unrecognizable when face to large
    scale bubble. Furthermore, in term of mask annotation, under the assumption
    that all bubble has a elliptical shape in 2D, so if some anomal form the
    masks because a coarse map than pixel-wise ground truth. For the overlap
    annoation, the labels will be assign approximated relatively base on visible
    part.
  personalSensitiveInformation: >-
    TiBuDB does not contain any personal or sensitive information. The images
    are from physical gas-liquid experiments and do not involve human subjects,
    identity, or private data.
  dataUseCases: >-
    TiBuDB is used for bubble deep learning vision-based detector which use for
    Bubble Size Distribution and Sauter Mean Diameter in gas-liquid mass
    transfer analysis
  dataSocialImpact: >-
    TiBuDB is mainly used for scientific research in fluid mechanics and
    chemical engineering. It can help improving understanding of gas-liquid mass
    transfer and industrial process efficiency. There is no direct negative
    social impact expected because no human-related data is involved. However,
    using this dataset outside of its intended experimental context may lead to
    incorrect interpretation or unreliable model performance.
task_categories:
  - object-detection
  - image-segmentation
language:
  - en
tags:
  - oriented-bounding-box
  - obb
  - multi-task
  - computer-vision
size_categories:
  - n<1K
configs:
  - config_name: yolo_obb
    data_dir: yolo_obb
    default: true
  - config_name: yolo_det
    data_dir: yolo_det
  - config_name: yolo_seg
    data_dir: yolo_seg
  - config_name: dota
    data_dir: dota
  - config_name: roboflow
    data_dir: roboflow
    data_files:
      - split: train
        path:
          - images/train/_annotations.coco.json
          - images/train/*.png
      - split: validation
        path:
          - images/valid/_annotations.coco.json
          - images/valid/*.png
  - config_name: coco_standard
    data_dir: coco
    data_files:
      - split: train
        path:
          - images/train/1.0_train_coco.json
          - images/train/*.png
      - split: validation
        path:
          - images/val/1.0_val_coco.json
          - images/val/*.png
  - config_name: coco_obb
    data_dir: coco
    data_files:
      - split: train
        path:
          - images/train/1.0_train_coco_obb.json
          - images/train/*.png
      - split: validation
        path:
          - images/val/1.0_val_coco_obb.json
          - images/val/*.png
  - config_name: large_image_test
    data_dir: large_image_test
    description: >-
      Metrological reference set. Includes raw images, binary masks, and ImageJ
      measurements (mm) for physical accuracy assessment.
dataset_info:
  features:
    - name: image
      dtype: image
    - name: label
      dtype: string

🫧 TiBuDB Dataset

TiBuDB is a multi-task benchmark for bubble flow detection. It supports 5+ annotation formats including HBB, OBB, and Segmentation.

🛠️ Toolkit & Scripts

To facilitate multi-format training, we provide a specialized toolkit in the /tools directory:

• Format Converters: obb2coco.py, yolo2coco.py, and yolo2dota.py for seamless transitions between annotation standards.
• Data Integrity: check_duplicated_yolo.py to ensure dataset uniqueness and correct_cocorf.py for Roboflow-specific COCO fixes.
• Quick Instance and Image Counts: count_instances.py to count images and instance quantity in train and val splits.
• Visualization: viz_samples.py,viz_yolo_det.py,viz_yolo_obb.py,viz_yolo_seg.py, viz_yolo_together.py,viz_dota.py scripts to overlay OBB and Segmentation masks for manual verification.

⚠️ Annotation Notes: Edge Cases

• Boundary Bubbles: Bubbles located at the image edges may appear as standard HBB (Horizontal Bounding Boxes) rather than OBB.
• Reason: Rotation angles are difficult to determine when the bubble is truncated by the frame.
• Impact: When training OBB models, these instances provide a 'neutral' rotation signal.

Dataset Structure

The dataset is organized into subfolders for each format. Select the configuration in the Hugging Face viewer to preview specific formats.