alakxender/dhivehi-image-bbox-prompt

Dhivehi Image Bounding Box Prompt Dataset

This dataset, alakxender/dhivehi-image-bbox-prompt, contains 58,738 images annotated with COCO-style bounding boxes and Dhivehi (Thaana script) text, along with layout categories such as Text, Title, Picture, Caption, and Columns. It is designed for OCR, document layout analysis, and multimodal vision–language research focused on Dhivehi.

Dataset

Each row includes:

• image — the RGB image (preserved original dimensions)

• width, height — image size in pixels

• image_id — unique integer id

• annotations — list of objects with:

  • bbox — bounding box (COCO-style [x, y, width, height], absolute pixels)
  • category — one of: Text, Title, Picture, Caption, Columns
  • text — Dhivehi transcript for the region (may be empty for pictures)

• num_objects — number of boxes in the image

• image_path — original file path (useful for local debugging)

Splits

Split	Count
train	52,864
validation (test)	5,874
total	58,738

Usage

from datasets import load_dataset

ds = load_dataset("alakxender/dhivehi-image-bbox-prompt")
sample = ds["train"][0]

print(sample["image_id"], sample["width"], sample["height"])
print(sample["annotations"][0])  # {'bbox':[x,y,w,h], 'category':'Text', 'text':'...'}

Validate bbox format quickly:

x, y, w, h = sample["annotations"][0]["bbox"]
assert w > 0 and h > 0  # COCO-style width/height

Target tasks & model ideas

• OCR (text detection + recognition)
• Document layout understanding
• Multimodal VLM

Sample

From assets/visualize_hf_dataset.py:

Category	Color (hex)
Text	#FF0000 (red)
Title	#00FF00 (green)
Picture	#0000FF (blue)
Caption	#FF00FF (magenta)
Columns	#FFFF00 (yellow)

Visualize a few samples (helper script)

# First 8 training images (streamed)
python visualize_hf_dataset.py --split train --head 8

# Skip first 500 validation images and take 12 (streamed)
python visualize_hf_dataset.py --split validation --offset 500 --head 12

# Random 10 from the first 2000 training items (streamed reservoir sampling)
python visualize_hf_dataset.py --split train --sample 10 --limit 2000

# Show Dhivehi text snippets and downscale to max 1600px side
python visualize_hf_dataset.py --split train --head 6 --show-text --max-side 1600

Key options

• --streaming / --no-streaming – Enable/disable HF streaming mode (default: streaming on)
• --head N – Take the first N samples (with optional --offset)
• --sample N --limit M – Randomly pick N samples from the first M streamed examples (reservoir sampling)
• --show-text – Overlay a short snippet of the Dhivehi text for each region
• --make-grid N – Make a simple grid collage of the saved visualizations

All generated images are saved in the specified --out-dir (default: hf_visualizations/).

Known limitations

• Some captions fields may be empty or contain line-break artifacts.
• Bounding boxes are axis-aligned; no polygons/rotations are provided.
• Class set is focused on page layout (not general object categories).
• Images are included in their original resolutions and orientations, resulting in varying sizes and aspect ratios.

Notes

This dataset was created for testing and experimentation in layout analysis of printed Dhivehi documents, and is not intended as a fixed standard; the dataset, annotation schema, or category set may be modified to suit specific project requirements.