remote-gesture

Remote Gesture

A small convolutional classifier that turns a single-stroke touch gesture into one of ten television remote actions. It runs in the browser via TensorFlow.js on the phone that drew the gesture; no inference happens server-side.

Files

Input

A 32×32 single-channel image scaled to [0, 1], produced by the rasterisation described in the dataset card. The greyscale ramp along the stroke encodes direction, so stroke order is part of the input, not just shape.

Architecture

Input(32, 32, 1)
Conv2D(32, 3, relu, padding=same) -> MaxPool2D(2)
Conv2D(64, 3, relu, padding=same) -> MaxPool2D(2)
Flatten -> Dense(128, relu) -> Dropout(0.4) -> Dense(10, softmax)

Training

• Optimiser: legacy Adam, clipnorm=1.0
• Learning rate: CosineDecayRestarts(1e-4, first_decay_steps=50, t_mul=2.0, m_mul=0.9, alpha=1e-7)
• Loss: sparse categorical cross-entropy
• Batch size 32, up to 500 epochs
• EarlyStopping(monitor='val_loss', patience=30, restore_best_weights=True, min_delta=1e-4)
• Trained on the dataset's train split and validated on test; augmentation touches only train, so validation is un-augmented real data
• Each training class is oversampled to 10,000 with RandomZoom((-0.1, 0.3)), RandomRotation(0.02), and RandomTranslation(0.1, 0.1), plus per-class flips and rotations gated by the allowMirrorHorizontal, allowMirrorVertical, allowRotation, and allowSlanted flags in gestures.json

Intended use

The application rejects any prediction below 0.6 confidence and treats it as "unknown" — it vibrates three times and sends nothing. That threshold is part of how the model is used and should be carried over by anyone reusing it; the model has no reject class of its own.

Limitations

• One author's handwriting, one device, one screen geometry.
• Sensitive to stroke direction by design.
• Ten fixed classes; adding one requires retraining and republishing both the dataset and the model.