ricmatsui/remote-gesture

Remote Gesture

Single-stroke touch gestures for a television remote, captured on a phone and rasterised to 32×32 greyscale images. Ten classes: directional arrows, a circle and counter-circle for power on/off, letter shapes for app shortcuts, and caret/vee for home and back.

How a sample is produced

Understanding the rasterisation matters, because the images are not raw drawings and cannot be reproduced without it.

1. Pointer positions are sampled during a drag, keeping a new point only once it is more than clientWidth / 100 from the previous one.
2. The stroke's bounding box is scaled to fit a 32×32 canvas with 2px padding, preserving aspect ratio, and centred.
3. It is drawn on black with lineWidth 2 and round caps and joins.
4. The stroke colour ramps from hsl(0, 0%, 10%) at the start to hsl(0, 0%, 90%) at the end. The greyscale gradient encodes stroke direction. This is what makes circle and counterCircle separable in a single static image, and it means the model is sensitive to stroke order, not only to shape.
5. Pixels are scaled to [0, 1] at training time.

Classes

Label order is significant and matches the model's output index:

up, down, left, right, circle, counterCircle, n, p, v, ^

Splits

A stratified 80/20 split, so both sides keep the class balance above. Training augments only the train split, leaving test as un-augmented real samples.

Limitations

• A single author, a single device, one screen geometry.
• No adversarial or accidental-input examples; the application handles those with a 0.6 confidence threshold rather than a reject class.
• Class counts are uneven, from 262 (p) to 388 (^); training oversamples each class to 10,000 with per-class augmentation.