---

Overview

AnnotateANU is a full-stack image annotation platform for computer vision datasets. It includes a SAM3 inference service, a core API for workflow and data management, and a React/Konva UI for annotation, exploration, and exports.

Features

• Annotation canvas with rectangle, polygon, and point tools, selection, vertex editing, zoom/pan, undo/redo, and shortcuts.
• SAM3 text and bbox prompts with single, auto-apply, and batch modes plus auto-detect helpers.
• Projects, tasks, and jobs with assignment, status tracking, approvals, and train/val/test splits.
• File share browser with uploads, shared image registry, project image pools, and tag/category metadata.
• Explore view with a virtualized gallery, filters, bulk tagging, and analytics panels.
• Analytics panels: Dataset statistics, annotation analysis, spatial heatmaps, and image quality metrics with multi-select filtering.
• Image quality metrics: Background processing for sharpness, brightness, contrast, uniqueness with real-time progress tracking.
• Exports: server-side datasets (COCO JSON, manifest CSV, image folders) plus local COCO/YOLO export from the annotation app.
• BYOM registry with inference proxy, plus admin user management and role-based auth.

Demo

Hero GIFs (Top Priority)

Demo	GIF	Description
Landing		Marketing landing and onboarding
Annotation workflow		Prompt, mask, edit, and export in the canvas

AI-Powered Annotation

Demo	GIF	Description
SAM3 Text Prompt		Type "cat" → instant segmentation mask
SAM3 Bbox Prompt		Draw rough box → SAM refines to perfect polygon
Polygon Simplification		Reduce polygon points for easier editing
BYOM Inference		Bring your own model for custom detection

Explore & Analytics

Demo	GIF	Description
Virtualized Gallery		Smooth scrolling through 1000+ images
Tri-State Filtering		Include/exclude/idle tag states with AND/OR
Histogram Filters		Click histogram bars to filter by dimension
Quality Metrics		Background processing with progress tracking
Confidence Slider		Filter annotations by confidence range

Note: Replace placeholder GIFs above with actual recordings. See GIF Recording Guide for details.

Why AnnotateANU?

Feature	AnnotateANU	CVAT	FiftyOne	Encord	Roboflow
Self-Hosted	✅	✅	✅	❌	⚠️
SAM3 Integration	✅ Native	✅	❌	✅	✅
Text Prompts	✅	❌	❌	✅	✅
BYOM Support	✅	⚠️	❌	⚠️	❌
Tri-State Tag Filters	✅	❌	⚠️	⚠️	❌
Histogram Filtering	✅	❌	✅*	⚠️	✅
Quality Metrics	✅	❌	⚠️*	✅	❌
Model Source Tracking	✅	❌	✅	⚠️	❌
Polygon Simplification	✅	❌	❌	❌	❌

*Requires Python code

Key Differentiators

• 🎯 Tri-State Tag Filtering: Include, exclude, or ignore tags with AND/OR logic—unique to AnnotateANU
• 📊 Integrated Analytics: Built-in histograms, heatmaps, and quality metrics without code
• 🤖 BYOM Registry: Register and use your own inference endpoints alongside SAM3
• ⚡ Real-time Confidence Filtering: Slider-based annotation filtering by confidence score
• 🔧 Polygon Simplification: Reduce polygon complexity for better performance

Architecture

flowchart LR
  user["Annotator"] --> web["Web UI (React + Vite)"]
  web --> core["API Core (FastAPI)"]

  core --> db[(PostgreSQL)]
  core --> redis[(Redis)]
  core --> share["File Share /data/share"]
  core --> exports["Export Storage /data/exports"]
  core --> worker["API Core Worker (Celery)"]
  worker --> exports

  core --> proxy["Inference Proxy"]
  proxy --> sam3["SAM3 Inference API"]
  sam3 --> model["SAM3 Model (HF Transformers)"]
  proxy -.-> byom["BYOM Endpoints"]

Diagram source: docs/architecture/system-overview.mmd.

Services (Docker dev stack)

Service	Responsibility	Host port	Container port
Web app	Annotation UI, dashboards, explore, exports	18712	5173
API Core	Auth, projects, tasks, jobs, model registry, exports	18711	8001
SAM3 Inference API	Text, bbox, batch segmentation	18710	8000
API Core Worker	Export jobs and background tasks	-	-
PostgreSQL	Core data store	5432	5432
Redis	Cache and task queue	-	6379

Host ports use the 18710-18712 block instead of the defaults (8000, 8001, 5173) to avoid clashing with other services on the same machine. Only the host side is remapped — inside the anu-network bridge the services still talk over their standard ports, so SAM3_API_URL=http://backend:8000 and friends stay unchanged.

To use a different block, edit the ports: entries in docker/docker-compose.yml and keep these in sync:

• apps/web/.env — VITE_SAM3_API_URL, VITE_CORE_API_URL (baked into the browser bundle, so they must be host-reachable URLs)
• apps/api-core/.env and the CORS_ORIGINS override in docker/docker-compose.yml — must list the web app's origin
• Makefile — the URLs echoed by the docker-up / docker-rebuild targets

Quick Start

Prerequisites

• Docker + Docker Compose
• HuggingFace token for facebook/sam3 (model access is gated)

Configure env files

cp apps/api-inference/.env.example apps/api-inference/.env
cp apps/api-core/.env.example apps/api-core/.env
cp apps/web/.env.example apps/web/.env

Add your HuggingFace token in apps/api-inference/.env:

HF_TOKEN=hf_your_token_here

Start the stack

make docker-up

Services:

• Web: http://localhost:18712
• SAM3 API docs: http://localhost:18710/docs
• API Core docs: http://localhost:18711/docs

Local Development (no Docker)

# SAM3 inference
make backend-install
make backend-run

# API core
make core-install
make core-run

# Web app
make frontend-install
make frontend-dev

These run on the application defaults — SAM3 on 8000, API Core on 8001, Vite on 5173 — not the remapped Docker host ports. apps/web/.env.example points at the Docker ports, so when running without Docker set VITE_SAM3_API_URL=http://localhost:8000 and VITE_CORE_API_URL=http://localhost:8001 in apps/web/.env.

Configuration

• apps/api-inference/.env
  • HF_TOKEN (required)
  • SAM3_MODEL_NAME, SAM3_DEVICE (auto, cpu, cuda)
  • MAX_IMAGE_SIZE_MB, MAX_BATCH_SIZE, MAX_IMAGE_DIMENSION
• apps/api-core/.env
  • DATABASE_URL, DATABASE_URL_SYNC
  • REDIS_URL
  • SAM3_API_URL
  • JWT_SECRET_KEY
• apps/web/.env
  • VITE_SAM3_API_URL
  • VITE_CORE_API_URL
  • VITE_ENV

API Docs

• SAM3 Inference API: http://localhost:18710/docs
• API Core: http://localhost:18711/docs

(Docker dev stack ports — see Services. Running locally without Docker, these are 8000 and 8001.)