superlabel

superlabel is a lightweight desktop annotation tool for object detection datasets. It is built with Tauri, SolidJS, Konva, and Rust, with a focus on fast manual labeling, local-first workflows, and ONNX-assisted pre-labeling.

Features

• Local image and label folder workflow.
• Workspace memory: reopen a recent project and resume from the last labeled image.
• Manual annotation tools:
  • Rectangle
  • Polygon
  • Point
  • Circle
  • Line
• Polygon snapping to existing vertices and edges.
• Resize rectangle annotations by dragging corners or edges.
• Canvas zoom and pan with mouse wheel.
• Class management:
  • Add classes
  • Rename classes
  • Switch selected annotation class from the annotation list
  • Switch class with number keys: 1-9, 0
• Undo and redo for annotation edits.
• Autosave toggle.
• Language switch: English and Chinese.
• Export format switch: YOLO and COCO.
• ONNX-assisted labeling:
  • Load a user-provided ONNX model
  • Download preset YOLOv8 ONNX models
  • Run pre-labeling on current image
  • Run pre-labeling on all images
  • Filter ONNX suggestions by class range
  • Show download and inference progress
• Project statistics panel:
  • Per-class sample distribution
  • Unannotated image count
  • Average bbox width, height, area, and aspect ratio
  • Aspect-ratio distribution
  • Estimated remaining labeling time

Data Formats

YOLO Labels

Rectangle annotations are saved as YOLO txt files in the selected label folder:

class_id cx cy width height

All values except class_id are normalized to [0, 1].

Example:

0 0.512000 0.438000 0.210000 0.184000

Shape Sidecar Files

Non-rectangle shapes are saved in sidecar files next to YOLO labels:

image_name.superlabel.json

This keeps YOLO compatibility while allowing superlabel to store polygons, points, circles, and lines.

Classes

Class names are saved in:

classes.txt

COCO Export

When output format is set to COCO, superlabel writes:

annotations.json

Current COCO export is based on rectangle annotations.

ONNX Models

superlabel supports YOLO-style ONNX detection outputs through the Rust backend.

Preset models include:

• YOLOv8n/s/m/l/x COCO detection
• YOLOv8n segmentation ONNX
• YOLOv8x OIV7
• YOLOv8s World v2

Notes:

• Segmentation mask decoding is not implemented yet.
• Custom text prompts for world/open-vocabulary models are not implemented yet.
• Preset downloads are restricted to the configured X-AnyLabeling release URLs.

Tech Stack

• Tauri 2
• Rust
• SolidJS
• Konva
• Vite
• tract-onnx
• image

Development

Install dependencies:

pnpm install

Run the Tauri app in development:

pnpm tauri dev

If the Linux desktop window crashes on Wayland with a Gdk-Message protocol error, run the X11 fallback:

pnpm tauri:dev:x11

Linux builds disable WebKitGTK DMABUF rendering and compositing by default to avoid blank windows on some Wayland, NVIDIA, AMD, and Intel driver combinations. Packaged builds also default to XWayland when they detect a Wayland session. To force the app through XWayland manually, run:

SUPERLABEL_FORCE_X11=1 superlabel

To keep native Wayland enabled on a machine where it is stable, run:

SUPERLABEL_USE_WAYLAND=1 superlabel

Run only the frontend dev server:

pnpm dev

Build the frontend:

pnpm build

Build the desktop app:

pnpm tauri build

Linux release builds generate AppImage and deb packages. Use the AppImage on Arch Linux and CachyOS.

Check Rust code:

cd src-tauri
cargo check

Check TypeScript:

npx tsc --noEmit

Project Structure

.
├── src
│   ├── components       # UI components and canvas interaction
│   ├── stores           # SolidJS app state
│   ├── utils            # file APIs, i18n, YOLO helpers, workspaces
│   └── types            # shared TypeScript types
├── src-tauri
│   └── src              # Rust backend commands
├── index.html
├── package.json
└── vite.config.ts

Rust Backend Responsibilities

The Rust backend handles the parts that benefit from native performance and file-system access:

• Reading image folders and label files
• Writing labels, classes, shape sidecars, and COCO exports
• Image dimension detection
• ONNX model download
• ONNX model caching and inference
• Project statistics calculation

Workspace Behavior

superlabel remembers recently opened projects locally. A workspace records:

• Image folder path
• Label folder path
• Last opened image
• Autosave setting
• Output format
• Language

Recent workspaces are stored in browser local storage inside the app runtime. They are not written as project files yet.

Current Limitations

• COCO export currently covers rectangle annotations.
• YOLO txt files only represent rectangle annotations.
• Segmentation mask decoding is not implemented.
• Open-vocabulary prompt configuration is not implemented.
• Workspace records are local app state, not portable workspace files.

Roadmap Ideas

• Portable .superlabel-workspace files.
• COCO export for polygons and other shapes.
• Segmentation mask decoding for YOLOv8-seg.
• Configurable ONNX model profiles with preprocessing/output schema presets.
• Dataset quality warnings for class imbalance, tiny boxes, and unusual aspect ratios.
• Keyboard shortcut customization.