Quick Draw

Notice: This project is under active development. Configuration values, default parameters, and documented behavior may not reflect the latest state of the code. Users are responsible for verifying all settings against the source files before use in production. Bug reports and corrections are welcome.

Real-time sketch recognition on the SolidRun RZ/V2N board.

Draw on a touchscreen and the DRP-AI3 accelerator classifies your sketch from 345 categories in ~1 ms. Single C++ binary — no server, no Python, no sockets at runtime.

Board	RZ/V2N SR SOM (SolidRun HummingBoard iIoT)
CPU	ARM Cortex-A55
Accelerator	DRP-AI3 (AI-MAC @ 1 GHz)
Display	1024 x 600 touchscreen (Wayland)
Model	MobileNetV2, 345 classes, INT8
Inference	~1 ms

Quick Start

# Build (from host — Docker cross-compiles for ARM64)
cd board_app
./docker_build.sh

# Deploy
./deploy.sh <board-ip>

# Run (on board)
ssh root@<board-ip>
cd /home/root/quickdraw && ./run.sh

Documentation

Document	Contents
docs/TRAINING.md	Dataset download, model training, ONNX export
docs/BUILD.md	DRP-AI compilation, C++ build, packaging, deployment
docs/APP.md	Application architecture, configuration, controls

Project Layout

quickdraw/
+-- train/
|   +-- download_ndjson.py       Download + render strokes at 128x128
|   +-- train.py                 Train MobileNetV2, export ONNX
|   +-- data_128/                345 .npy files
|
+-- calibration/                 1,725 PNG images for INT8 quantization
+-- drpai_model/                 Compiled DRP-AI model
+-- best_model.pt                Trained weights (14 MB)
+-- qd_model.onnx                ONNX model (14 MB)
+-- categories.txt               345 class names
+-- generate_calibration.py      Calibration image generator
|
+-- board_app/
|   +-- src/                     C++ source (gui, inference, preprocessing)
|   +-- config.ini               DRP-AI frequencies and model path
|   +-- config.json              UI layout, colors, AI comments
|   +-- labels.txt               345 class names
|   +-- docker_build.sh          Build from host via Docker
|   +-- package.sh               Create deploy/ folder
|   +-- deploy.sh                SCP to board
|   +-- compile_model.sh         DRP-AI TVM compilation
|   +-- run.sh                   Board startup script
|   +-- lib/                     MERA2 runtime libraries
|   +-- deploy/                  Ready-to-copy package (46 MB)
|
+-- docs/
    +-- TRAINING.md
    +-- BUILD.md
    +-- APP.md