AutoLabel UI

English | 简体中文

AutoLabel UI is an unofficial desktop auto-labeling client based on NVIDIA LocateAnything. It can run either against a user-owned Linux GPU server over SSH or against a local LocateAnything runtime prepared on the user's own computer.

The project focuses on the desktop UI, job orchestration, media preparation, YOLO conversion, result visualization, and server/local runtime adapters. It does not include model weights, official NVIDIA inference runtime source code, Docker images, datasets, credentials, logs, or private deployment material.

AutoLabel UI is a personal open-source project by Huang Jiawen. It is not affiliated with, sponsored by, or endorsed by NVIDIA.

Screenshots

Clean screenshots use empty/demo UI states only.

What It Does

• Select images, folders, or videos from a Windows desktop UI.
• Configure open-vocabulary class prompts for NVIDIA LocateAnything.
• Run jobs in either remote SSH mode or local runtime mode.
• Export YOLO images, labels, data.yaml, raw detections, visualized boxes, and sampled annotated videos.
• Track task progress, GPU telemetry, results, and local task history.
• Keep SSH passwords in process memory only; saved config excludes passwords.

Runtime Modes

Remote Server Mode

Use this mode when you have a Linux GPU server with Docker and an existing LocateAnything runtime. AutoLabel UI uploads each job over SSH/SFTP, starts the container task, monitors status files, downloads results.zip, verifies its SHA256, and extracts results locally.

Local Runtime Mode

Use this mode when you want to run without a server. In the settings page, choose Local runtime, then set:

• the Python executable from your LocateAnything runtime environment, such as a .venv\Scripts\python.exe;
• the local nvidia/LocateAnything-3B model/runtime directory that contains the official batch_infer.py;
• an optional local workspace directory for runtime locks and temporary jobs.

The repository still does not ship model weights or official runtime files. The user must download them separately and comply with NVIDIA's model license.

Quick Start

Requirements:

• Windows 10/11 x64
• Python 3.11 or newer

git clone https://github.com/mybrain-naozi/AutoLabel-UI.git
cd AutoLabel-UI
install_dependencies.bat
run_autolabel.bat

Manual setup:

py -3 -m venv .venv
.venv\Scripts\python.exe -m pip install -r requirements.txt
.venv\Scripts\python.exe run_autolabel.py

Application data is stored in %LOCALAPPDATA%\AutoLabel by default. Set AUTOLABEL_DATA_ROOT to use a different local data directory.

Remote Server Contract

AutoLabel UI does not provide a server, Docker image, model, or inference runtime. For remote mode, prepare:

1. A Linux server with SSH/SFTP access and an NVIDIA GPU.
2. A Docker container that the SSH user can start and execute commands in.
3. A host job root mounted into the container workspace.
4. A .venv Python environment inside the container workspace.
5. A user-downloaded nvidia/LocateAnything-3B directory containing the official batch_infer.py runtime files.
6. Runtime dependencies required by the official inference stack, such as PyTorch, Transformers, Pillow, and OpenCV.

Default example values:

Setting	Example
Host job root	/srv/autolabel/locateanything
Container workspace	/workspace/locateanything
Container name	locateanything
Model directory	/workspace/locateanything/models/LocateAnything-3B
Hugging Face endpoint	https://huggingface.co

Jobs are created under <host job root>/jobs/<job id> and mapped to <container workspace>/jobs/<job id> in the container.

Security Notes

• Do not commit screenshots, logs, configs, media, labels, or task outputs that reveal private projects, datasets, hostnames, accounts, paths, or credentials.
• SSH passwords are not saved to config files, logs, or job files.
• Verify SSH host fingerprints before enabling unknown-host acceptance.
• Use a least-privilege SSH account and a dedicated server job root.

License Boundary

The AutoLabel UI source code in this repository is released under Apache License 2.0.

The NVIDIA Eagle / LocateAnything code is released under Apache-2.0, but the nvidia/LocateAnything-3B model uses a separate NVIDIA License. At the time this notice was written, that model license limits the model and derivative works to non-commercial research or evaluation use. Always check the official license before using or redistributing anything derived from the model.

Official sources:

• Model repository and license: https://huggingface.co/nvidia/LocateAnything-3B
• Official Eagle code: https://github.com/NVlabs/Eagle/tree/main/Embodied

See MODEL_LICENSE_NOTICE.md for details.