Children's Drawing Explorers

Interactive web pages for exploring a large dataset of children's drawings (48 categories, ages 2–10).

🔗 Landing page: https://vislearnlab.github.io/drawing-explorer/

The site root (index.html) is a landing page linking the explorers below; it's the single URL to cite. The two in-repo explorers are:

• explorer.html — the CLIP-embedding map of all ~37,770 drawings, colored by model recognizability (described in What it shows, below).
• strokes.html — a stroke-by-stroke player: drawings render from their per-stroke SVG paths, colored by agreed semantic part, with autoplay and a side-by-side compare mode. Data is built by build_strokes.py into strokes_data/. See build_strokes.py's docstring for the source inputs.

What the CLIP map (explorer.html) shows

• Drawings laid out by a 2-D t-SNE of the 48-d CLIP per-category probability vectors. Each point is one drawing.
• Color by: CLIP recognizability (probability of the intended category, default), correct/incorrect, log-odds, age, or category.
• Filters: category, age range, minimum recognizability, correct-only.
• Hover a point to see the drawing + its scores; click to pin it.
• The footer stat bar updates with the count / % CLIP-correct / mean recognizability of the currently visible subset.

Run locally

python3 -m http.server 8000
# open http://127.0.0.1:8000/

(open http://127.0.0.1:8000/ for the landing page; drawing PNGs live in drawings/.)

Files

• index.html — landing page linking the explorers (cite this URL).
• explorer.html — the self-contained CLIP-map explorer (no build step).
• strokes.html — the self-contained stroke-by-stroke player (no build step).
• points.json — t-SNE layout + per-drawing CLIP scores consumed by explorer.html.
• strokes_data/ — per-category stroke JSON consumed by strokes.html.
• drawings/ — the 37,770 drawing PNGs (150×150).
• build_data.py — regenerates points.json from the source CLIP embeddings and recognizability tables (see below).
• build_strokes.py — regenerates strokes_data/ from the agreed part labels and raw per-stroke SVGs.

License

Licensed under CC BY-NC 4.0 (free to use with attribution, non-commercial). See LICENSE. Please credit the Visual Learning Lab (UC San Diego) and cite the dataset paper (above).

Rebuilding points.json

build_data.py expects the source data from the drawing_production_and_recognition repository (CLIP feature .npy, metadata, and merged_clip_class_and_meta.csv). Point the paths at that checkout and run:

python3 build_data.py

Data & paper

Drawings and recognizability scores come from:

Long, B. et al. Parallel developmental changes in children's production and recognition of line drawings of visual concepts. PsyArXiv · OSF

License

Drawings and data: CC BY-NC-SA 4.0 (per the source dataset).