mohamedemam/words-detection

Word Detection — multi-source YOLO dataset

A single-class object-detection dataset for locating word bounding boxes in scanned documents, screenshots, handwritten notes and photographed pages. Four public Roboflow datasets were merged into one YOLO-format release with a single word class.

This dataset is the detection stage of an OCR pipeline: it finds words, it does not read them. Recognition is a separate model.

At a glance

Task	Object detection (bounding boxes)
Classes	1 — word
Images	7,003
Boxes	574,576
Annotation format	YOLO (class cx cy w h, normalized)
Image format	JPEG
License	CC BY 4.0 (attribution required)
Sources	4 Roboflow projects

Splits

Split	Images	Boxes	Boxes / image (mean · median · max)
train	6,337	520,866	82.2 · 74 · 360
val	422	38,505	91.2 · 76 · 334
test	244	15,205	62.3 · 47 · 250
total	7,003	574,576	82.1 · 74 · 360

Splits come from the upstream projects and were preserved, so images never move between train, val and test. 80 duplicate images were dropped during the merge (all from math_words).

Sources and attribution

Every upstream project is CC BY 4.0. Attribution is required if you use this dataset.

Source	Roboflow project	Original classes	Images	Boxes
devanagari_words	aaa-8kwq1/devanagari-word-detection	word	297	23,662
word_detection_yolo	yogesh-jangir-3utla/word-detection-using-yolo	word	458	86,454
words_5tolc	aabis-siyas-hoisi/words-5tolc	Texto	1,031	14,075
math_words	kyiv-national-university/math-words	Text, W, math, text	5,217	450,385

math_words is distributed through the mirror el-emam/math-words-oahhj, because the original project has no generated export version.

What the merge changed

• All classes collapsed to word (id 0). Texto, Text, W, math and text were treated as the same concept. No boxes were dropped for class reasons.
• Split names normalized: valid → val, test kept.
• Duplicate images removed (80, all from math_words), identified across sources.
• Coordinates validated: every label line has exactly 5 fields, class id 0, and coordinates inside [0, 1] with positive width and height.
• Filenames namespaced as <source>__<original-name>.jpg so the origin of any image is recoverable.

Structure

.
├── data.yaml            # Ultralytics dataset config (nc: 1, names: [word])
├── classes.json         # class list and id mapping
├── sources.json         # provenance and per-source counts
├── stats.json           # full statistics, machine readable
├── manifest.json        # merge manifest produced by the builder
├── README.md
├── images/
│   ├── train/  (6,337)
│   ├── val/    (422)
│   └── test/   (244)
└── labels/
    ├── train/  (6,337 .txt)
    ├── val/    (422 .txt)
    └── test/   (244 .txt)

Each label file matches an image by basename and contains one line per word:

class_id  center_x  center_y  width  height     # all normalized to [0, 1]

Geometry

Property	Min	Median	Max
Image width (px)	640	640	4,032
Image height (px)	328	640	4,032
Box width (normalized)	—	0.060	0.987
Box height (normalized)	—	0.037	0.914

The median word box is 0.060 × 0.037 of the image, i.e. roughly 39 × 24 pixels on a 640 px image. This is a small-object detection problem: training at 640 px will merge or lose many boxes. 960–1280 px is recommended.

Median box aspect ratio is 1.54 (p95 = 5.5), so words are typically wider than tall, but long compound words reach ratios above 5.

Usage

Ultralytics YOLO

from ultralytics import YOLO

model = YOLO("yolo26s.pt")
model.train(
    data="data.yaml",
    imgsz=960,
    batch=16,
    epochs=40,
    single_cls=True,
    max_det=512,      # up to 360 boxes per image
)
metrics = model.val(split="test")

max_det matters here: some images contain more than 300 words, so the Ultralytics default of 300 truncates predictions and understates mAP.

Loading the images and labels

from pathlib import Path

from PIL import Image

root = Path(".")  # dataset root
for label in (root / "labels" / "val").glob("*.txt"):
    image = Image.open(root / "images" / "val" / f"{label.stem}.jpg")
    for line in label.read_text().splitlines():
        cls, cx, cy, w, h = map(float, line.split())
        box = [
            (cx - w / 2) * image.width,
            (cy - h / 2) * image.height,
            (cx + w / 2) * image.width,
            (cy + h / 2) * image.height,
        ]

Limitations

• One class only. No distinction between printed and handwritten text, no script labels, no reading order, no transcription.
• Script imbalance. devanagari_words contains Devanagari text, while the other three are mostly Latin and mathematical notation. Boxes are correct, but the dataset is not script-balanced and should not be described as Latin-only.
• Domain imbalance. math_words contributes 78% of all boxes; mathematical documents dominate the label distribution. Plain prose pages are under-represented.
• Small boxes. Median word height is ~3.7% of the image. Downstream users should train at high resolution and validate on their own domain.
• No negative images. Every image contains at least one word box, so the model sees no explicit background-only examples.
• Inherited label noise. Box quality is that of the upstream projects; no manual re-annotation was performed.

License

Released under CC BY 4.0, inherited from the four upstream datasets. You must credit:

• aaa-8kwq1 — Devanagari word detection
• yogesh-jangir-3utla — Word Detection using yolo
• aabis-siyas-hoisi — Words
• Kyiv National University — math-words

Citation

@misc{gemma_ocr_words_real,
  title  = {Word Detection: multi-source YOLO dataset},
  author = {Emam},
  year   = {2026},
  note   = {Merged from four CC BY 4.0 Roboflow datasets},
  url    = {https://huggingface.co/datasets/elemam/gemma-ocr-words-real}
}

Provenance

Field	Value
Version	1.0.0
Generated	2026-11-10
Merge seed	42
Builder	scripts/build_real_word_dataset.py
Statistics	scripts/dataset_stats.py