workshed/ipad-ui-sketches

Hand-Drawn iPad UI Sketches

133 low-fidelity sketches of mobile app screens, drawn with an Apple Pencil on an iPad, with bounding-box labels for 7 UI component classes (1,265 boxes in total).

The dataset was collected for the MSc dissertation Towards More Efficient Mobile UI Design: Automatic Code Generation from Hand-Drawn Sketches Using Deep Learning (Daniel Leivers, University of Bath, 2023). It was used to train a YOLOv5 model that runs on-device in an iPad app and turns the user's drawing into SwiftUI code as they draw.

• Code, trained model and dissertation PDF: https://github.com/Otaku-Development/sketch-to-swiftui
• Demo video: https://www.youtube.com/watch?v=SKGdZ3H9eyY

Dataset structure

The files are laid out in the standard Ultralytics YOLO format. Each image folder also has a metadata.jsonl, which lets the Hugging Face datasets library and dataset viewer load the boxes.

images/{train,validation,test}/<uuid>.jpg      # the sketch
images/{train,validation,test}/metadata.jsonl  # boxes for 🤗 datasets (COCO-style [x_min, y_min, width, height], pixels)
labels/{train,validation,test}/<uuid>.txt      # YOLO labels: class x_center y_center width height (normalised 0–1)
data.yaml                                      # Ultralytics dataset config
build_dataset.py                               # script that produced the splits from the raw export

Images are black strokes on a white background. 132 images are 913–914 × 1860 px and one is 1016 × 2070 px; all have a portrait iPhone aspect ratio.

Splits

Split	Images	button	image	label	navigationBar	tabBar	textField	switchButton
train	107	288	162	457	38	38	29	15
validation	13	46	22	52	6	5	16	2
test	13	29	12	41	2	1	1	3
total	133	363	196	550	46	44	46	20

The splits are a seeded random 80/10/10 split, using the first seed from 42 upwards (45) that puts every class in every split. They are not the split used in the dissertation, which was not kept, so results trained on these splits won't match the dissertation's numbers exactly.

Classes

ID	Class
0	button
1	image
2	label
3	navigationBar
4	tabBar
5	textField
6	switchButton

These map to the SwiftUI components the app generates. In the drawings, an image is usually a rectangle with a cross through it and a label is usually one or more wavy lines.

Usage

Ultralytics YOLO

from huggingface_hub import snapshot_download
from ultralytics import YOLO

path = snapshot_download("workshed/ipad-ui-sketches", repo_type="dataset")
YOLO("yolov8s.pt").train(data=f"{path}/data.yaml", imgsz=640, epochs=300)

🤗 Datasets

from datasets import load_dataset

ds = load_dataset("workshed/ipad-ui-sketches")
ds["train"][0]["objects"]  # {'bbox': [[x, y, w, h], ...], 'category': ['image', 'label', ...]}

Collection process

The sketches were collected with a purpose-built iPad data-capture app (source code in the GitHub repository above). Each volunteer was shown a screenshot of an existing iPhone app and asked to draw a low-fidelity version of it. The app suggested bounding boxes from the pen strokes; the volunteer then confirmed, merged or split them and labelled each one. Sketches and labels were uploaded to Firebase and later exported to YOLO format.

• Contributors: 10 volunteers.
• Device: iPad with Apple Pencil. There are no paper or scanned sketches.
• Collected: 2023.

The reference app screenshots shown to volunteers are not included. Only the volunteers' own drawings are.

Cleaning applied for this release

• One sketch had been saved 14 times, with identical image and labels. The 13 duplicate copies were removed (146 → 133 images).
• The raw export had both a .png and a .jpg of each sketch, with identical content. Only the .jpg is kept.

Personal information and consent

Contributors agreed to a consent screen saying their sketches and labels would be used to train a classification model. The files are named with random UUIDs and contain only pen strokes and box coordinates. No names, contributor IDs or device information are included. Each contributor's optional name was stored separately and is not part of this dataset.

If you contributed a sketch and want it removed, please open an issue on the GitHub repository or start a discussion on this dataset page.

Known limitations

• Small. 133 images from 10 people. It works best for fine-tuning a pre-trained detector, as a seed for a larger collection, or as an evaluation set, not for training from scratch.
• Class imbalance. label, button and image make up 88% of boxes. switchButton has 20 examples and the test split has only 1–3 examples of some classes, so per-class test metrics for the rare classes are very noisy.
• segmentedControl removed. Volunteers labelled 7 segmented controls, but this class was dropped from the export. The images that contain them are still included, so those segmented controls appear without a box and a model will learn to treat them as background.
• No contributor IDs. Splits are random by image, not by person, so the same person's drawing style can appear in both train and test. This makes test scores optimistic for drawers the model hasn't seen.
• One drawing style per device. Every sketch is a clean single-weight digital stroke on white. Expect a domain gap if you apply a model trained on this data to paper sketches or photos.
• Annotation quality. Each sketch was labelled once, by the person who drew it, with no second review.

Licence

Creative Commons Attribution 4.0 (CC BY 4.0). Please cite the dissertation if you use the dataset.

Citation

@mastersthesis{leivers2023sketch,
  author = {Leivers, Daniel},
  title  = {Towards More Efficient Mobile {UI} Design: Automatic Code Generation from Hand-Drawn Sketches Using Deep Learning},
  school = {University of Bath},
  type   = {{MSc} dissertation},
  year   = {2023},
  url    = {https://github.com/Otaku-Development/sketch-to-swiftui}
}