Annotix

Open-source desktop platform for ML dataset annotation, training, and collaboration
Images · Video · Audio · Time Series · Tabular Data

Download  •  Read the Paper  •  System Overview  •  Build from Source  •  Cite

---

Paper

Annotix: An Open-Source Desktop Platform for Comprehensive Machine Learning Dataset Annotation

Published on Preprints.org (April 2026)

Universidad Austral de Chile, Campus Puerto Montt — TecMedHub

If you use Annotix in your research, please cite the paper.

---

Why Annotix?

Most annotation tools focus on a single data type or require cloud accounts. Annotix is different:

---

Status

Last updated: June 2026 — v2.9.10

:white_check_mark: = production-ready  ·  :construction: = usable but incomplete  ·  :warning: = known limitation

---

Download

Pre-built binaries for the latest release:

Linux packaging is a self-contained tarball (annotix-v<version>-linux-x86_64.tar.gz) with the binary, libpdfium.so and a run.sh launcher — no .deb/.rpm/AppImage. It needs libwebkit2gtk-4.1-0, libgtk-3-0 and librsvg2-2 on the system, and can register itself as a desktop entry. Windows installers bundle FFmpeg DLLs and pdfium.

All releases: github.com/Debaq/Annotix/releases

The app checks GitHub for new releases and shows an in-app update banner with a dynamic changelog.

---

Features at a Glance

Annotation Tools

7 tools on a high-performance Konva canvas:

Plus: mouse wheel zoom, image rotation, label/grid toggles, per-annotation visibility, draggable floating panels persisted per project, quick class selection (1-0, Q-P for up to 20 classes), and undo/redo with 100-step history.

SAM-Assisted Segmentation

Segment Anything runs locally via ONNX — no cloud, no API key.

• AMG mode: a 16×16 point grid (32×32 in HQ mode) feeds the decoder in batch; click a candidate mask + press a class key to convert it into the active tool's format (BBox / OBB / Mask / Polygon)
• Refine mode: click-by-click positive/negative prompts on a cached image embedding
• Each candidate keeps all 3 multimask outputs as uint8 logits at 256px, so the granularity slider switches between them without re-running the decoder
• Frontend sliders (granularity, score, NMS, overlap) re-filter candidates without re-running AMG
• Encoder/decoder models are stored app-level ({data_dir}/sam_models/), shared across projects
• Candidates are ephemeral — never written to project.json

Project Types

23 project types across four families, each with its own editor and export path:

• Images (9) — detection, oriented detection, semantic segmentation, instance segmentation, polygon, keypoints, landmarks, single- and multi-label classification
• Time Series (9) — classification, forecasting, anomaly detection, segmentation, pattern recognition, event detection, regression, clustering, imputation
• Audio (4) — classification, speech recognition, sound event detection, TTS recording
• Tabular (1) — built-in editor with column selection and scikit-learn training

Video is a capture path rather than a project type: FFmpeg extracts frames at a configurable rate into the project's images, annotated as tracks with keyframes and linear interpolation, then baked to per-frame annotations.

Images can also be ingested from PDF documents (pages rasterized natively with pdfium) and stored as WebP per project.

Integrated ML Training (19 Backends)

Train models directly from the app with real-time metrics charts, a training monitor with suggestions ("coach"), free-text observations, and an exportable PDF training report.

Full backend list

Object Detection

Backend	Models
YOLO (Ultralytics)	YOLO26, YOLOv8–v12
RT-DETR (Ultralytics)	RT-DETR-l, RT-DETR-x
RF-DETR (Roboflow)	RF-DETR-base, RF-DETR-large
MMDetection (OpenMMLab)	30+ architectures (Faster R-CNN, DINO, Co-DETR, etc.)

Semantic Segmentation

Backend	Models
SMP	U-Net, DeepLabV3+, FPN, PSPNet, etc.
HuggingFace Segmentation	SegFormer, Mask2Former, etc.
MMSegmentation	Full OpenMMLab catalog

Instance Segmentation

Backend	Models
Detectron2 (Meta)	Mask R-CNN, Cascade R-CNN, etc.

Keypoints & Pose

Backend	Models
MMPose	HRNet, ViTPose, RTMPose, etc.

Oriented Object Detection (OBB)

Backend	Models
MMRotate	Oriented R-CNN, RoI Transformer, etc.

Image Classification

Backend	Models
timm	700+ models (ResNet, EfficientNet, ViT, ConvNeXt, etc.)
HuggingFace Classification	ViT, BEiT, DeiT, Swin, etc.

Time Series

Backend	Task
tsai	Classification, regression, forecasting
PyTorch Forecasting	TFT, N-BEATS, etc.
PyOD	Anomaly detection
tslearn	Temporal clustering
PyPOTS	Missing value imputation
STUMPY	Matrix Profile (motif/pattern discovery)

Tabular

Backend	Task
scikit-learn	RandomForest, SVM, kNN, GradientBoosting, etc.

4 execution modes:

6 training presets: small_objects, industrial, traffic, edge_mobile, medical, aerial.

Model export: PyTorch .pt, ONNX, TorchScript, TFLite, CoreML, TensorRT.

Inference

Run trained or third-party ONNX models over a whole project:

• Auto-detection of architecture and metadata (classes, input size, nc/names) across five output families: YOLOv8+, YOLOv5–v7 (objectness), YOLOv10/YOLO26 (end-to-end, no NMS), multi-output (SSD, EfficientDet, Faster R-CNN, DETR) and classification
• Batch inference with cancel, per-prediction accept/reject, conversion to annotations
• Model archives can be dropped in directly (ZIP extraction + drag & drop import)
• Execution providers are opt-in (TensorRT, CUDA, DirectML, CoreML); CPU path uses SIMD preprocessing and a parallel pipeline

P2P Collaboration

Real-time collaborative annotation powered by Iroh (QUIC). No central server, no port forwarding: N0 discovery (pkarr + DNS + relays) keeps peers reachable across networks and re-syncs automatically when a peer comes back online.

• Host or join with a session code; the host secret is encrypted at rest
• Roles: LeadResearcher (full control) / Annotator / DataCurator, with per-session rule overrides (upload, export, edit classes, delete) and live role changes
• Live sync of images, annotations and classes with author attribution; blob transfer with retries and resumable downloads
• Image presence locks (3-min TTL, auto-renewed) so you can see who is on which image
• Work distribution: auto-assign batches across peers, adjust assignments, per-peer progress stats
• Data approval queue: submissions from collaborators can require host approval
• Pause / resume a session; peer list with heartbeat-based online status

Conflict resolution is per-key last-write-wins (iroh-docs). Video, audio and time-series assets are not part of the synced document — share those via TIX export.

Network Sharing (Serve)

Publish a project over the LAN as an HTTP server with a bearer-token protected web UI — collaborators annotate from a browser with no install. It picks a free port, enumerates local addresses and self-checks reachability before announcing the URL. Optional auto-save.

REST endpoints: /api/projects, /api/projects/{id}, /api/projects/{id}/images, and per-image …/{image_id}, …/file, …/thumbnail, …/annotations (GET and POST).

Browser Automation

Train on Google Colab for free (T4 GPU) via Chrome DevTools Protocol:

• Auto-detects Chromium browsers, uploads dataset, runs training
• Real-time progress with pause / resume / cancel

Query LLMs without API keys through the user's browser: Kimi, Qwen, DeepSeek, HuggingChat.

Export & Import

17 export formats: YOLO Detection, YOLO Segmentation, COCO JSON, Pascal VOC, CSV (Detection / Classification / Keypoints / Landmarks), Folders by Class, U-Net Masks, TIX (native), rasterized preview (with and without labels), HuggingFace ASR, LJSpeech, CSV Audio Classification, CSV Sound Events.

8 import formats with automatic detection: YOLO (detection & segmentation), COCO, Pascal VOC, CSV (4 variants), U-Net Masks, Folders by Class, TIX. Multiple .tix files can be merged, homogenizing class sets.

Inspector & Filters

Annotation inspector reachable from the project gallery, with debug/observation filters over marks, classes and gallery, per-class counters, and multi-project comparison.

Keyboard Shortcuts

All shortcuts are fully customizable from Settings with per-context conflict detection.

Default shortcuts

Image Tools

Shortcut	Action
B	Bounding Box
O	OBB
M	Mask
P	Polygon
K	Keypoints
L	Landmarks
V	Select
H	Pan
[ / ]	Decrease / Increase brush size
E	Toggle eraser
A / D	Rotate image
Enter	Confirm drawing
Esc	Cancel drawing

SAM assist is toggled from the canvas toolbar; while active, Tab cycles candidates, Esc exits refine mode and the class keys accept the hovered mask.

Navigation

Shortcut	Action
PageUp / PageDown	Previous / Next sample
Ctrl++ / Ctrl+-	Zoom in / out
Ctrl+0	Zoom to fit

General

Shortcut	Action
Ctrl+S	Save
Ctrl+Z / Ctrl+Y	Undo / Redo
Del / Backspace	Delete selection
Esc	Deselect

Quick Class Selection

Keys	Classes
1 - 0	Classes 1 to 10
Q - P	Classes 11 to 20

Video

Shortcut	Action
T	New track

Time Series

Shortcut	Action
V	Select
P	Point
R	Range
E	Event
A	Anomaly

Audio

Shortcut	Action
F2	Play / Pause
F3 / F4	Replay / Rewind
Left / Right	Scrub
Enter	Split
Tab	Save & next

TTS Recording

Shortcut	Action
Space	Record
Enter	Accept take
R	Repeat
S	Skip sentence

Languages

10 languages across 47 namespaces, with lazy loading and English fallback:

de Deutsch · en English · es Espanol · fr Francais · it Italiano · ja Japanese · ko Korean · pt Portugues · ru Russian · zh Chinese

---

Architecture

+-----------------------------------------------------+
|                    Frontend                           |
|   React 19 + TypeScript + Tailwind + shadcn/ui       |
|   Konva (canvas) . Chart.js (metrics) . i18next      |
|   Zustand (state) . React Router 7                   |
+-----------------------------------------------------+
|                  Tauri 2 IPC                          |
|             194 registered commands                   |
+-----------------------------------------------------+
|                  Backend (Rust)                       |
|   +------------+ +-----------+ +-----------------+   |
|   |   Store    | | Commands  | | Export/Import   |   |
|   | (JSON+RAM) | | (24 mod)  | | (17+8 formats)  |   |
|   +------------+ +-----------+ +-----------------+   |
|   +------------+ +-----------+ +-----------------+   |
|   |  Training  | | Browser   | | P2P (Iroh)      |   |
|   | (19 backs) | | Automat.  | | QUIC mesh       |   |
|   +------------+ +-----------+ +-----------------+   |
|   +------------+ +-----------+ +-----------------+   |
|   | Inference  | |    SAM    | | Serve (axum)    |   |
|   | (ONNX/ort) | | enc/dec   | | LAN web UI      |   |
|   +------------+ +-----------+ +-----------------+   |
+-----------------------------------------------------+
|               External Integrations                   |
|   Python (micromamba) . FFmpeg . pdfium . Chromium    |
|   Cloud APIs . Iroh P2P network                      |
+-----------------------------------------------------+

Storage

All data stored as JSON + raw assets on disk. No database.

~/.local/share/annotix/config.json        -> global configuration (projects_dir)
~/.local/share/annotix/sam_models/        -> SAM encoder/decoder ONNX (app-level)
~/.local/share/annotix/p2p/iroh/          -> P2P blobs + replicated docs
{projects_dir}/{uuid}/project.json        -> the whole project
{projects_dir}/{uuid}/images/             -> original images
{projects_dir}/{uuid}/thumbnails/         -> generated thumbnails
{projects_dir}/{uuid}/videos/             -> video files
{projects_dir}/{uuid}/audio/              -> audio files
{projects_dir}/{uuid}/models/             -> registered inference models
{projects_dir}/{uuid}/training/           -> per-job datasets and results

project.json is a single document holding metadata, classes, images with their annotations, time series, videos with their tracks, audio, tabular data, training jobs with per-epoch metric history, inference models, TTS sentences and the collaboration config. An in-memory cache with dirty-flag tracking sits in front of it; writes go to a .tmp file and are then renamed.

IDs are UUID v4 end to end. Class IDs are reindexed to their position on save, and the remap propagates to image annotations, video tracks, audio and time series.

---

Tech Stack

Frontend

Technology	Version	Purpose
React	19	UI framework
TypeScript	5.7	Static typing
Vite	6	Bundler and dev server
Tailwind CSS	3.4	Utility-first styling
shadcn/ui	—	Component library (Radix UI)
Zustand	5	Global state with persistence
React Router	7	SPA routing
Konva	10	2D annotation canvas
Chart.js	4	Metrics visualization
jsPDF / html2canvas	4 / 1.4	Training report PDF
TanStack Virtual	3	Virtualized gallery
i18next	24	Internationalization

Backend (Rust)

Crate	Version	Purpose
tauri	2	Desktop application framework
serde / serde_json	1	JSON serialization
image / imageproc	0.25	Image processing
fast_image_resize	5	SIMD resizing
rayon	1.10	Data parallelism
geo	0.29	Polygon simplification
webp	0.3	WebP encoding
ort	2.0-rc	ONNX Runtime (inference + SAM)
ffmpeg-the-third	4.1	Video frame extraction
pdfium-render	0.8	PDF page rasterization
zip	2	Export/import packaging
quick-xml	0.37	Pascal VOC XML
csv	1.3	CSV import/export
reqwest	0.12	HTTP client (cloud providers)
jsonwebtoken	9	GCP service-account auth
axum	0.8	LAN serve HTTP server
chacha20poly1305 / subtle	0.10 / 2	Session secret encryption
headless_chrome	1.0	Browser automation (CDP)
iroh + blobs/gossip/docs	0.96–0.98	P2P networking (QUIC)
tokio	1	Async runtime
blake3	1	Hashing

Python (via micromamba)

Package	Purpose
ultralytics	YOLO, RT-DETR
rfdetr	RF-DETR
mmdet, mmseg, mmpose, mmrotate	OpenMMLab suite
segmentation-models-pytorch	Semantic segmentation
timm	Classification (700+ models)
detectron2	Instance segmentation
tsai, pytorch-forecasting	Time series deep learning
pyod, tslearn, pypots, stumpy	Time series classical ML
scikit-learn	Tabular ML

---

System Requirements

• OS: Windows 10+, macOS 12+, Linux (glibc 2.31+)
• RAM: 4 GB minimum, 8 GB recommended
• Disk: ~500 MB for the app + space for datasets
• GPU (optional): NVIDIA with CUDA or Apple Silicon with MPS for accelerated training
• FFmpeg: required for video annotation (bundled in release builds)
• Chromium browser (optional): for browser automation (Chrome, Brave, Edge)

---

Getting Started

Prerequisites

• Node.js >= 18 (CI uses Node 24)
• pnpm (the repo ships pnpm-lock.yaml)
• Rust >= 1.89
• Tauri 2 prerequisites for your platform

Build & Run

git clone https://github.com/Debaq/Annotix.git
cd Annotix
pnpm install
pnpm tauri:dev       # development (hot-reload)
pnpm tauri:build     # production build

Scripts

CI runs cargo fmt/clippy, cargo nextest and the frontend build on every push.

---

Project Structure

annotix/
├── src/                         # React frontend (~46k LOC)
│   ├── App.tsx                  # Router and providers
│   ├── lib/
│   │   ├── db.ts                # Type definitions (mirrors Rust structs)
│   │   ├── tauriDb.ts           # Centralized Tauri invoke bridge
│   │   └── i18n.ts              # i18next configuration
│   ├── components/ui/           # shadcn/ui components
│   └── features/
│       ├── canvas/              # Annotation canvas (7 tools)
│       ├── sam/                 # SAM assist (overlay, panel, store)
│       ├── gallery/             # Virtualized gallery + filters
│       ├── video/               # Video annotation
│       ├── audio/               # Audio annotation + TTS recording
│       ├── timeseries/          # Time series annotation
│       ├── tabular/             # Tabular data editor
│       ├── classification/      # Classification workflow
│       ├── training/            # ML training panel + monitor + PDF report
│       ├── export/              # 17 export formats
│       ├── import/              # 8 import formats
│       ├── inference/           # Model inference
│       ├── p2p/                 # P2P collaboration
│       ├── serve/               # LAN sharing dialog
│       ├── browser-automation/  # Chrome automation
│       ├── setup/               # First-run setup
│       └── settings/            # App settings
├── src-tauri/                   # Rust backend (~41k LOC)
│   └── src/
│       ├── lib.rs               # 194 Tauri command registrations
│       ├── store/               # Storage layer (state, IO, cache)
│       ├── commands/            # 24 command modules
│       ├── export/              # Export format modules
│       ├── import/              # Import + auto-detector + merge
│       ├── training/            # Multi-backend ML pipeline + cloud providers
│       ├── browser_automation/  # Headless Chrome (Colab, LLM chat)
│       ├── p2p/                 # Iroh P2P networking
│       ├── serve/               # axum LAN server + web UI
│       └── inference/           # ONNX inference + SAM (encoder/decoder/AMG)
├── docs/                        # Roadmaps and backend references
├── wiki/                        # User documentation
└── public/locales/              # 10 languages x 47 namespaces

---

Citation

If you use Annotix in your research, please cite:

@article{annotix2026,
  title     = {Annotix: An Open-Source Desktop Platform for Comprehensive Machine Learning Dataset Annotation},
  year      = {2026},
  publisher = {Preprints.org},
  url       = {https://www.preprints.org/manuscript/202604.0919}
}

Full paper: https://www.preprints.org/manuscript/202604.0919

---

Contributing

Contributions are welcome. Please open an issue first to discuss what you'd like to change.

---

License

MIT License — TecMedHub, Universidad Austral de Chile, Campus Puerto Montt.