English | 日本語

dolla_one_recognizer

A Python 3 implementation of the $1 Unistroke Recognizer.

詳しい日本語の解説は README.ja.md にあります。

What is this

The $1 Unistroke Recognizer is a lightweight single-stroke gesture recognition algorithm by Wobbrock, Wilson, and Li (UIST '07). It needs no machine-learning training: register just one template per shape and it works. It resamples the input stroke to a fixed number of points, normalizes rotation, scale, and position, then picks the nearest registered template.

Input is a plain list of points [[x, y], ...], so the coordinates can come from a mouse, touch input, or a tracker such as OpenCV optical flow / color / fingertip tracking — just append each frame's (x, y) and pass the list in.

This repository contains the recognizer itself, packaged as dolla_one_recognizer_py/, plus two small Tkinter demo apps: export_gesture_app.py (draw and save gesture templates) and recognize_app.py (match a drawn stroke against saved templates).

Install

git clone https://github.com/yuzujelly2222/dolla_one_recognizer_python.git
cd dolla_one_recognizer
pip install -r requirements.txt

Dependencies are numpy (geometry) and pillow (PNG I/O for the demo apps).

Usage

from dolla_one_recognizer_py import dolla_one_recognizer

# Register templates — one example stroke per shape is enough
line_template = [[x, x] for x in range(0, 101, 5)]
check_template = [[0, 30], [30, 0], [90, 60]]

recognizer = dolla_one_recognizer(
    size=250,
    templates=[line_template, check_template],
    templates_name=["line", "checkmark"],
    n=64,
)

# Raw points to recognize (e.g. collected from a mouse drag)
query = [[x, x * 0.9 + 2] for x in range(0, 101, 4)]

best_points, best_name, score = recognizer.recognize(query)
print(best_name, score)  # -> line 0.96...

Try the GUI demos:

python export_gesture_app.py   # draw and save to gestures/ as CSV + PNG
python recognize_app.py        # auto-load gestures/ and test recognition

How it works

$1 applies four steps to the input points and returns the nearest template:

1. Resample (_resample) to n (default 64) evenly spaced points, absorbing differences in drawing speed.
2. Rotate (_rotate_to_zero) so the angle from the centroid to the first point is zero.
3. Scale & translate (_scale_to_square -> _translate_to_origin) into a size x size square (default 250) centered at the origin.
4. Match (recognize) against each template, searching +/-45 degrees for the rotation that minimizes the average point-to-point distance (Golden Section Search), and return the closest. The distance is converted to a 0-1 score.

The Protractor enhancement (a closed-form optimal angle instead of the search) is not included in this implementation. See README.ja.md for a fuller walkthrough.

API

Only the members below are part of the public API; the geometry helpers are prefixed with _ and are internal.

License & citation

Distributed under the New BSD License — see LICENSE.

This is an independent implementation based on the pseudocode from the paper below. The original authors and copyright holders exist separately from this port.

Wobbrock, J.O., Wilson, A.D. and Li, Y. (2007). Gestures without libraries, toolkits or training: A $1 recognizer for user interface prototypes. Proceedings of the ACM Symposium on User Interface Software and Technology (UIST '07), pp. 159-168.

Official site (algorithm, pseudocode, reference implementation): https://depts.washington.edu/acelab/proj/dollar/index.html