> For the complete documentation index, see [llms.txt](https://dev.solid-run.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://dev.solid-run.com/renesas/rz-v2n/rz-v2n-other-articles/ai-demo-quick-draw-doodle-recognition-rz-v2n.md). # AI Demo: Quick Draw Doodle Recognition (RZ/V2N) {% embed url="" %} Live recognition on the HummingBoard-IIoT โ€” top-5 guesses re-rank with every stroke. {% endembed %} ### ๐Ÿ“‹ Overview **Production-grade AI, on a single industrial SOM, at a price built for scale.** **You draw. The board knows. In one millisecond โ€” locally, from 345 categories.** No GPU server. No cloud subscription. No per-inference billing. Just one **industrial SOM โ€” the RZ/V2N with the DRP-AI3 accelerator** โ€” handling everything on-device. You draw on the touchscreen. The board **recognises what you are sketching** โ€” live, in milliseconds. A **MobileNetV2** classifier, quantised to **INT8** and accelerated by **DRP-AI3 (15 Sparse / 4 Dense TOPS)**, picks the right answer from **345 categories** in roughly **one millisecond**. #### See it in action **Sketch a circle.** The model offers its best guesses โ€” *cookie ยท pizza ยท sun ยท clock*. **Add two pointed ears on top.** The top prediction pivots โ€” *cat*. **Add whiskers and a tail.** Confidence locks in โ€” *cat, 92 %.* That is the demo: live recognition that **sharpens with every stroke**. Draw an apple, a bicycle, an airplane, a house, a guitar, a hamburger โ€” any of the **345 categories** in Google's Quick, Draw! taxonomy โ€” and the board's **5 best guesses re-rank live**. ๐Ÿ‘‰ Source: [**github.com/SolidRun/quickdraw-rzv2n**](https://github.com/SolidRun/quickdraw-rzv2n) *** ### ๐ŸŽฏ Why It Matters **Real-time AI at the touch surface.** Sub-millisecond inference. No perceptible latency between stroke and prediction. **Cloud-free by design.** Drawings, predictions, and feedback never leave the board. A natural fit for privacy-sensitive HMI, kiosks, education, and industrial controls. **Cost-effective edge intelligence.** DRP-AI3 delivers GPU-class inference performance at a small fraction of the price. No discrete GPU. No add-on accelerator card. No per-inference cloud bill. Just one industrial SOM doing the work โ€” at a price point that makes mass deployment realistic. **A complete, reproducible recipe.** Train โ†’ quantise โ†’ compile โ†’ cross-compile โ†’ deploy โ†’ run. Fully scripted. Fully documented. Use it as the starting point for your own edge-AI product.
*** ### ๐Ÿง  The Platform The demo runs on the **HummingBoard-IIoT** carrier with the **RZ/V2N SOM** โ€” SolidRun's industrial Edge-AI platform. The headlines are below. For the full datasheets and pinouts, follow the links beneath each table. {% columns %} {% column %} **RZ/V2N SOM โ€” the brain** | Highlight | Value | | --------------- | ----------------------------------------------------------------- | | **AI engine** | **DRP-AI3 โ€” 15 Sparse TOPS / 4 Dense TOPS** | | **CPU** | 4ร— Arm Cortex-A55 @ up to **1.8 GHz** + 1ร— Cortex-M33 (real-time) | | **Memory** | Up to **8 GB LPDDR4 with inline ECC** | | **Form factor** | **47 ร— 30 mm**, industrial **โˆ’40 ยฐC to +85 ยฐC** | โ†’ Full specs: [RZ/V2N SOM product page](https://www.solid-run.com/embedded-industrial-iot/renesas-rz-family/rz-v2n-som/) ยท [Hardware User Manual](https://dev.solid-run.com/renesas/rz-v2n/com-som/rz-v2n-som-hardware-user-manual) {% endcolumn %} {% column %} **HummingBoard-IIoT โ€” the carrier** | Highlight | Value | | -------------------- | --------------------------------------------------------- | | **Industrial I/O** | RS232/RS485 ยท CAN-FD ยท 2ร— GbE ยท USB 3.2 ยท **PoE 802.3at** | | **Display + camera** | MIPI-DSI ยท MIPI-CSI 4-lane | | **Power** | 7โ€“32 V wide-input, reverse-polarity protected, or PoE | | **Operating + OS** | Industrial **โˆ’40 ยฐC to +85 ยฐC** ยท Yocto Linux + Weston | โ†’ Full specs: [HummingBoard RZ/V2N IIOT SBC product page](https://www.solid-run.com/embedded-industrial-iot/renesas-rz-family/hummingboard-rz-series-sbcs/hummingboard-rz-v2n-iiot-sbc/) ยท [Quick Start Guide](https://dev.solid-run.com/renesas/rz-v2n/sbc-platform/hummingboard-iiot-rz-v2n-som-quick-start-guide) {% endcolumn %} {% endcolumns %} {% hint style="info" %} The demo uses a **1024 ร— 600 touchscreen** over MIPI-DSI, running under **Weston (Wayland)**. No camera required โ€” input comes from the touch surface only. {% endhint %} *** ### ๐Ÿ—๏ธ How It Works A tight, deterministic loop between Cairo, the C++ preprocessor, the MERA2 runtime on DRP-AI3, and GTK3. ``` Touch / mouse input โ”‚ โ–ผ Cairo drawing surface (white background, black ink) โ”‚ โ–ผ C++ preprocessing โ€ข detect ink bounding box โ€ข crop, pad to square, invert โ€ข resize to 128 ร— 128, replicate to 3 channels โ€ข normalise to 0โ€“1 float32 โ”‚ โ–ผ DRP-AI3 inference (~1 ms) MobileNetV2 ยท INT8 ยท 345 classes โ”‚ โ–ผ Softmax โ†’ temporal smoothing โ†’ top-5 โ”‚ โ–ผ Display: prediction ยท confidence bars ยท AI commentary ``` The on-board preprocessor and the training pipeline share an **identical** normalisation convention. `mean = [0,0,0]`, `std = [1,1,1]` โ€” sketch-specific, **not** ImageNet. Mismatch these values and predictions will silently break โ€” [`TRAINING.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/TRAINING.md) flags this as **critical**. The [Configuration](#-configuration-in-brief) section repeats the warning below. Full reference: ๐Ÿ”— [`docs/APP.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/APP.md). *** ### ๐Ÿงฎ The Model | Property | Value | | ---------------- | ------------------------------------------------------------------------------------------- | | **Architecture** | MobileNetV2 โ€” ImageNet-pretrained backbone, custom 1280 โ†’ 768 โ†’ 345 head | | **Dataset** | [Google Quick, Draw!](https://github.com/googlecreativelab/quickdraw-dataset) โ€” 345 classes | | **Input** | `[1, 3, 128, 128]` (NCHW), grayscale replicated to 3 channels | | **Quantisation** | INT8 โ€” Percentile 99.99 calibration over 1,725 representative PNGs | | **Compiler** | DRP-AI Translator i8 v1.11 โ†’ DRP-AI TVM v2.7+ (MERA2 backend) | | **Inference** | **\~1 ms** on DRP-AI3 | | **Training** | Two-stage transfer learning โ€” \~6 hours on RTX 5060 Ti โ€” **\~82 % validation accuracy** | Training is **optional**. The repository ships a trained `qd_model.onnx` and a pre-compiled `drpai_model/` so you can deploy in minutes. Retrain only when you want to narrow the category set, swap the dataset, or experiment with the head. **Hardware requirement** โ€” training and building have different needs. {% tabs %} {% tab title="Train the model" %} A consumer GPU is enough. Reference run: **\~6 hours on an RTX 5060 Ti**, \~3 GB VRAM, \~82 % accuracy. You will need either: * A local **machine with an NVIDIA GPU** โ€” CUDA-capable, โ‰ฅ 4 GB VRAM at the default batch size, **or** * A **cloud ML training platform** โ€” Roboflow, Google Colab, Vertex AI, AWS SageMaker, Azure ML, or Paperspace โ€” that handles the GPU for you. {% endtab %} {% tab title="Build & deploy" %} **CPU-only.** No GPU needed. Runs on **any PC** with Docker: * Linux * macOS * Windows (WSL2) The Docker image carries the entire Renesas toolchain โ€” you do not install anything on the host beyond Docker itself. {% endtab %} {% endtabs %} For the full training recipe โ€” NDJSON download, augmentations, two-stage hyper-parameters, ONNX export verification, calibration generation โ€” see ๐Ÿ”— [`docs/TRAINING.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/TRAINING.md). ***
### ๐Ÿ› ๏ธ Build & Deploy {% hint style="info" %} **Just want to see it run?** Build the **`quickdraw-demo-image`** Yocto target instead โ€” a Weston-based image that auto-starts the demo on boot. ```bash MACHINE=rzv2n-sr-som bitbake quickdraw-demo-image ``` โ†’ See [**all available image targets in SolidRun's BSP layer**](https://github.com/SolidRun/meta-solidrun-arm-rzg2lc/tree/scarthgap_rzv2n_dev#available-image-targets) (branch `scarthgap_rzv2n_dev`). The pipeline below is the **development path** โ€” for retraining, recompiling, or modifying the app. {% endhint %} End to end, the pipeline is five stages โ€” most of them one command. {% stepper %} {% step %} **Toolchain โ€” one-time** Build the **DRP-AI TVM Docker image**. Bundles the Renesas RZ/V LP SDK, the DRP-AI Translator i8, and DRP-AI TVM (MERA2) โ€” reusable across every project. โ†’ [`docs/INSTALL.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/INSTALL.md) {% endstep %} {% step %} **Compile the model** `compile_model.sh` runs **inside the Docker container** โ€” quantises ONNX to INT8, runs the DRP-AI Translator, and emits a MERA2-ready runtime. โ†’ ONNX โ†’ INT8 โ†’ DRP-AI runtime {% endstep %} {% step %} **Cross-compile the C++ app** `docker_build.sh` runs **on the host**. It auto-detects the container, sources the SDK, and produces an ARM64 binary in `board_app/build/`. โ†’ `app_quickdraw_gui` (\~13 MB) {% endstep %} {% step %} **Package** Automatic at the end of `docker_build.sh`. Assembles a **\~46 MB `deploy/` folder** with the binary, model, MERA2 libs, configs, and `run.sh`. โ†’ Ready to ship. {% endstep %} {% step %} **Deploy & run** `deploy.sh --run` SCPs the bundle to the board and launches the app under **Weston (Wayland)**. First run installs MERA2 libs into `/usr/lib64/`. โ†’ Live in < 30 seconds. {% endstep %} {% endstepper %} The fast path: {% code title="host shell" %} ```bash # One-time: build the DRP-AI TVM Docker image (see INSTALL.md) cd ~/renesas/rzv_drp-ai_tvm docker build -t drp-ai_tvm_v2n_image_$(whoami) --build-arg PRODUCT=V2N . # Per-iteration: cross-compile, package, deploy, launch git clone https://github.com/SolidRun/quickdraw-rzv2n.git cd quickdraw-rzv2n/board_app ./docker_build.sh # cross-compile C++ + auto-package ./deploy.sh --run # scp + launch on the target ``` {% endcode %} That is the entire developer loop on a known-good toolchain. The scripts handle SDK sourcing, Docker container detection, the Wayland environment, and MERA2 library installation on first run. To recompile the model from a fresh ONNX export โ€” for example after retraining on a custom subset of categories โ€” the model-compile step lives inside the Docker container and is invoked via `docker cp` + `docker exec`. The exact recipe is documented step by step in ๐Ÿ”— [`docs/BUILD.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/BUILD.md). *** ### ๐Ÿ“š The Full Documentation Set The upstream repository's [`docs/`](https://github.com/SolidRun/quickdraw-rzv2n/tree/main/docs) folder is the authoritative, command-by-command reference. Use this page for orientation. Then follow the link that matches your task. | Goal | Read this | | ------------------------------------------------------- | -------------------------------------------------------------------------------------------- | | Set up the host PC, the Docker image, and the board OS | [`docs/INSTALL.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/INSTALL.md) | | Train the MobileNetV2 model and export ONNX | [`docs/TRAINING.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/TRAINING.md) | | Compile the model for DRP-AI, cross-compile, and deploy | [`docs/BUILD.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/BUILD.md) | | Understand the C++ application and its configuration | [`docs/APP.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/APP.md) | *** ### ๐ŸŽ›๏ธ Configuration in Brief Two text files in the deployed `deploy/` directory control everything tunable. Both are editable on the target. No rebuild required. * **`config.ini`** โ€” model path ยท DRP core frequency ยท AI-MAC frequency. * **`config.json`** โ€” UI layout ยท colour palette ยท prediction smoothing ยท confidence thresholds ยท the **AI commentary** system that drives the witty one-liners above the canvas. | What you can change without rebuilding | Where | | -------------------------------------------------- | ------------------------------------------------------------------- | | DRP / AI-MAC clock โ€” trade power vs. throughput | `config.ini` โ†’ `[drpai]` | | Canvas size, brush radius, refresh rate | `config.json` โ†’ UI keys | | Auto-predict delay, live-prediction interval | `config.json` โ†’ `auto_predict_delay_ms`, `live_predict_interval_ms` | | Confidence threshold and temporal smoothing window | `config.json` โ†’ `confidence_threshold`, `smooth_window` | | AI commentary tone, templates, no-repeat buffer | `config.json` โ†’ comment pools | | Brand text, logo placement | `config.json` โ†’ `title` / `subtitle` / `badge` | {% hint style="warning" %} **Normalisation must match across the pipeline.** Training, the DRP-AI compile patches, and the on-board C++ preprocessor must all use `mean = [0,0,0]` / `std = [1,1,1]`. A mismatch produces a working model that returns nearly random predictions โ€” the most common cause of "looks fine, classifies wrong." {% endhint %} Full reference: ๐Ÿ”— [`docs/APP.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/APP.md). *** ### ๐Ÿ“š References * [SolidRun โ€“ RZ/V2N SOM](https://www.solid-run.com/embedded-industrial-iot/renesas-rz-family/rz-v2n-som/) ยท [HummingBoard RZ/V2N IIOT SBC](https://www.solid-run.com/embedded-industrial-iot/renesas-rz-family/hummingboard-rz-series-sbcs/hummingboard-rz-v2n-iiot-sbc/) * [SolidRun โ€“ HummingBoard-IIoT & RZ/V2N SOM Quick Start Guide](https://dev.solid-run.com/renesas/rz-v2n/sbc-platform/hummingboard-iiot-rz-v2n-som-quick-start-guide) * [SolidRun โ€“ RZ/V2N SOM Hardware User Manual](https://dev.solid-run.com/renesas/rz-v2n/com-som/rz-v2n-som-hardware-user-manual) * [SolidRun GitHub โ€“ quickdraw-rzv2n](https://github.com/SolidRun/quickdraw-rzv2n) โ€” full source, scripts, and docs * [`docs/INSTALL.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/INSTALL.md) ยท [`docs/TRAINING.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/TRAINING.md) ยท [`docs/BUILD.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/BUILD.md) ยท [`docs/APP.md`](https://github.com/SolidRun/quickdraw-rzv2n/blob/main/docs/APP.md) * [Renesas โ€“ RZ/V2N AI Software Development Kit](https://www.renesas.com/software-tool/rzv2n-ai-software-development-kit) ยท [DRP-AI Translator i8](https://www.renesas.com/software-tool/drp-ai-translator-i8) * [Renesas DRP-AI TVM (rzv\_drp-ai\_tvm)](https://github.com/renesas-rz/rzv_drp-ai_tvm) ยท [DRP-AI Driver](https://github.com/renesas-rz/rzv2n_drp-ai_driver) * [Google Creative Lab โ€“ Quick, Draw! Dataset](https://github.com/googlecreativelab/quickdraw-dataset) * [PyTorch](https://pytorch.org/) ยท [ONNX](https://onnx.ai/) ยท [GTK 3 Reference Manual](https://docs.gtk.org/gtk3/) --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter: ``` GET https://dev.solid-run.com/renesas/rz-v2n/rz-v2n-other-articles/ai-demo-quick-draw-doodle-recognition-rz-v2n.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.