human_mouse

Replay real human mouse trajectories from the SapiMouse dataset onto a Playwright page.

Useful when you need mouse movement that is statistically indistinguishable from a real user — for ML-based bot-detection research, behavioral biometrics prototyping, or replay-based test fixtures.

• Source: https://git.vercanti.com/yellowdog/human_mouse
• License: MIT

Install

The package is not on PyPI yet. Install directly from the git remote:

# core only
uv add "human_mouse @ git+https://git.vercanti.com/yellowdog/human_mouse.git"

# with the bundled CloakBrowser example
uv add "human_mouse[demo] @ git+https://git.vercanti.com/yellowdog/human_mouse.git"

Or pin to a specific commit / tag:

uv add "human_mouse @ git+https://git.vercanti.com/yellowdog/human_mouse.git@v0.1.0"

pip install works with the same git+https://... URL.

One-time dataset download

import human_mouse as hm
hm.download_sapimouse("./data")   # ~8 MB, no registration

This unzips SapiMouse to ./data/sapimouse/userN/....

Quick start

import human_mouse as hm
from playwright.sync_api import sync_playwright

with sync_playwright() as pw:
    browser = pw.chromium.launch(headless=False)
    page = browser.new_page()
    page.goto("https://example.com")

    # one-shot: pick a real human trajectory matching the distance and replay it
    seg = hm.replay_random(
        page,
        start=(100, 100),
        end=(900, 500),
        data_root="./data/sapimouse",
        density=4,        # 4x spatial upsample for smoother visuals
        seed=42,
    )
    print(f"replayed {seg.user}/{seg.session}, {len(seg.points)} source points")

Step-by-step API

For full control over which segment is picked and how it's processed:

import human_mouse as hm

segs = hm.load_all_segments("./data/sapimouse")

seg = hm.pick_segments(
    segs,
    n=1,
    target_distance=1000,     # only match segments ~1000 px long
    distance_tolerance=0.3,   # ±30 %
    max_path_ratio=2.0,       # skip erratic meanderers
    seed=42,
)[0]

points = hm.affine_warp(seg, (100, 100), (900, 500))
points = hm.upsample(points, factor=4)
hm.replay(page, points, speed=1.0)

Public API

Symbol	Purpose
Segment	dataclass with .points, .start, .end, .straight_distance, .path_length, .path_ratio, .duration_ms
load_all_segments(data_root)	scan SapiMouse, return every continuous Move-only segment ≥ 50 pts
pick_segments(segments, n, ...)	filter by distance / path-ratio / distinct-session, randomly choose n
affine_warp(seg, start, end)	translate + rotate + uniform-scale onto new endpoints (shape preserved)
upsample(points, factor)	linearly interpolate factor-1 sub-points between every adjacent pair
replay(page, points, speed=1.0)	drive page.mouse.move(...) honoring the recorded dt
replay_random(page, start, end, data_root, ...)	one-shot: pick + warp + upsample + replay
download_sapimouse(dest)	download and unzip the dataset

Running the bundled example

The examples/ folder ships a Playwright canvas page (demo.html) plus a runner script that uses CloakBrowser. You need a local checkout for this since examples/ ships only in the sdist:

git clone https://git.vercanti.com/yellowdog/human_mouse.git
cd human_mouse
uv sync --extra demo
python -m human_mouse download ./sapimouse_data            # one-time, ~8 MB
uv run python examples/cloakbrowser_demo.py sapi           # one trajectory
uv run python examples/cloakbrowser_demo.py multi --n 10   # 10 overlaid

Outputs land in ./outputs/:

• sapi.png + sapi.json — single trajectory screenshot and per-event coords
• multi/overlay.png + per-run JSONs + summary.json

Development

git clone https://git.vercanti.com/yellowdog/human_mouse.git
cd human_mouse
uv sync --all-extras           # core + [demo] + [dev]
uv run pytest                  # 23 tests, ~40s with SapiMouse present
uv build                       # produces wheel + sdist in dist/

Integration tests auto-skip when sapimouse_data/sapimouse/ is missing, so the core 10 unit tests run on a fresh clone without the dataset.

What replay is and isn't

Is: a deterministic replay of one real human's mouse path, warped to your endpoints. The shape (curvature, hesitation, end-point fumbling) and the timing (dt distribution) come straight from a real recording.

Isn't: a generator. It samples from a fixed dataset of ~4 000 segments. For per-call novelty, randomize the seed; for true synthesis, look at trajectory generative models (e.g. SapiAgent, DMTG).

Dataset attribution

If you use SapiMouse data in published work, cite:

Antal, M. et al. SapiMouse: Mouse Dynamics-based User Authentication Using Deep Feature Learning. IEEE SACI 2021.

License

MIT.