AdaptiveFilter

A real-time adaptive Neural Network filter plugin for OpenTabletDriver that learns your movement patterns to provide smooth, low-latency cursor prediction.

Important

Please read the Disclaimer & Fair Play Notice before using this plugin.

⚠️Note: This project was vibe coded using AI assistance. The entire codebase, from the neural network implementation to the real-time web visualization, was developed through AI programming.

Features

• Neural Network Prediction: Learns your movement patterns in real-time using a Multi-Layer Perceptron
• Real-time Visualization: Web-based UI showing live predictions and network structure
• Configurable Architecture: Adjust hidden layers, learning rate, and network complexity
• Hybrid Mode: Choose between pure NN prediction or hybrid real input + NN upsampling
• Anti-Chatter Filter: Suppress noise and reduce jitter
• OneEuroFilter: Additional smoothing for stable predictions
• Training Metrics: Monitor accuracy, output rate, and training iterations

Installation

1. Download or Build:

   • Download the latest release from Releases
   • Or build from source: dotnet build -c Release

2. Install Plugin:

   • Open OpenTabletDriver
   • Open Plugin Directory
   • Copy AdaptiveFilter.dll to this directory
   • Restart OpenTabletDriver

3. Enable Filter:

   • In OpenTabletDriver, go to the Filters tab
   • Select AdaptiveFilter from the list
   • Click Apply

Configuration

Basic Settings

• Target Rate: Output refresh rate in Hz (default: 1000)
• Prediction Offset: Time offset for predictions in ms (default: 0)
• Lookahead: Multiplier for the prediction vector (default: 1.0)
• Web UI Port: Port for web interface (default: 5000)

Web UI

Access the visualization at http://localhost:5000

• Left panel: Real-time cursor visualization (Cyan = Input, Pink = Prediction)
• Right panel: Live neural network structure with weighted connections
• Stats: Prediction offset, output rate, accuracy (mm), and training iterations

Recommended Configurations

For Simple/Linear Movements

• 1 layer × 12 neurons (~360 parameters)
• Lower learning rate (0.005)

For Complex Aim Patterns (Default)

• 2 layers × 16 neurons (~2,080 parameters)
• Standard learning rate (0.01)

For Maximum Smoothness

• 3 layers × 12 neurons (~2,000 parameters)
• Higher learning rate (0.015)

Troubleshooting

Web UI not loading

• Ensure port 5000 is available
• Change Web UI Port in plugin settings

Cursor stops moving

• Reduce Hidden Layer Count to 1
• Lower Hidden Layer Size to 12

High latency

• Reduce Target Rate
• Enable Use Hybrid Mode

Jittery predictions

• Increase Samples count
• Increase Anti-Chatter Strength

How It Works

1. Input Processing: Tablet reports are filtered through Anti-Chatter filter
2. Training: Neural network learns movement deltas from sliding window of samples
3. Prediction: Network predicts next position based on recent movement patterns
4. Smoothing: OneEuroFilter reduces jitter while maintaining responsiveness
5. Upsampling: Predictions are emitted at configured target rate

Building from Source

git clone https://github.com/yourusername/AdaptiveFilter.git
cd AdaptiveFilter
dotnet build -c Release

Output: bin/Release/net6.0/AdaptiveFilter.dll

Dependencies

• .NET 6.0
• OpenTabletDriver.Plugin
• MathNet.Numerics

License

MIT License - See LICENSE file for details

Acknowledgments

Built with inspiration from various prediction and smoothing techniques used in tablet driver filters.