xtreme1-io/xtreme1

Xtreme1 is an all-in-one data labeling and annotation platform for multimodal data training and supports 3D LiDAR point cloud, image, and LLM.

Xtreme1 – Open‑source platform for multimodal training‑data annotation & management

What it is – Xtreme1 is a self‑hosted web application that lets you label, curate, and organise data for computer‑vision and large‑language‑model projects. It bundles a UI, back‑end services (MySQL, Redis, MinIO), and optional pre‑training inference containers so you can do pre‑labeling with built‑in models and then correct the results.

Why it matters – High‑quality annotated data is the bottleneck for vision and LLM pipelines. Xtreme1 provides a single place to handle 2‑D image boxes, polygons, key‑points, 3‑D LiDAR point‑cloud boxes, and sensor‑fusion annotations, plus ontology (class hierarchy) management and quality‑monitoring tools. The platform is open source (Apache‑2.0) and can be run locally or on a server via Docker, making it usable for research labs, startups, or enterprises that need full control over their data.

Core capabilities (as listed in the README)

Area	Features
Image annotation	Bounding boxes, polygons, polylines, key‑points. Uses YOLOR for optional auto‑labeling.
3‑D / sensor‑fusion annotation	LiDAR‑camera fusion boxes, powered by OpenPCDet and AB3DMOT.
Ontology centre	Define class hierarchies and attribute schemas that are exported to training pipelines.
Data management	Versioned storage (MinIO), database‑backed project metadata, quality‑monitoring dashboards.
Error detection	Tools to locate and fix labeling mistakes.
Result visualisation	Overlay model predictions on raw data for quick evaluation.
LLM‑focused RLHF tool	Beta‑stage interface for Reinforcement‑Learning‑from‑Human‑Feedback annotation loops.
Curation helpers	Visual debugging with MobileNetV3 embeddings and openTSNE dimensionality reduction.

Quick start (Docker‑compose)

Download the latest release (v0.9.3 at time of writing) and unzip.

wget https://github.com/xtreme1-io/xtreme1/releases/download/v0.9.3/xtreme1-v0.9.3.zip
unzip -d xtreme1-v0.9.3 xtreme1-v0.9.3.zip

Launch the stack
```
cd xtreme1-v0.9.3
docker compose up   # add -d to run in background
```
The compose file pulls MySQL, Redis, MinIO, plus the Xtreme1 backend and frontend containers. Data lives in Docker volumes, so it survives container restarts.
Open the UI – visit http://localhost:8190 (Chrome recommended).
Enable built‑in AI models (optional, requires an NVIDIA GPU with CUDA and the NVIDIA Container Toolkit):
```
docker compose --profile model up
```
This starts extra containers that serve YOLOR, OpenPCDet, etc., which the UI can call for pre‑labeling.

Installing from source (for developers)

Clone the repo, enable Docker BuildKit, and either build the images locally or keep the pre‑built images from Docker Hub.
The docker-compose.yml can be overridden with a docker-compose.develop.yml (already git‑ignored) to avoid committing local changes.
The README provides detailed steps for handling ARM CPUs, upgrading between releases, and backing up/restoring the MySQL database.

Typical use‑cases

Computer‑vision dataset creation – label 2‑D images or 3‑D point clouds for object detection, segmentation, or sensor‑fusion models.
LLM fine‑tuning with RLHF – use the beta RLHF annotation UI to collect human preference data for language‑model alignment.
Data quality pipelines – monitor annotation consistency, spot outliers, and re‑label problematic samples before training.
Team collaboration – multi‑user projects with role‑based access, shared ontologies, and versioned storage.

License & community

License: Apache License 2.0 (permissive, commercial‑friendly).
Governance: Hosted under the LF AI & Data Foundation; the project is positioned as the first open‑source labeling/annotation platform in that umbrella.
Support: Public issue tracker, documentation site (docs.xtreme1.io), and community channels on Twitter and Medium.

Bottom line – Xtreme1 is a full‑stack, Docker‑based solution for creating high‑quality multimodal training data. It is genuinely a software project aimed at AI/ML practitioners, not just a curated list or tutorial collection.