ArgusTrack

ArgusTrack is a annotation tool for multi-camera setups. It shows all camera views on one screen, supports 3D bounding boxes projected to a bird's-eye view (BEV), and saves annotations in a shared format under your dataset root.

Use it when you need consistent object IDs across cameras and frames, with calibration-aware placement on the ground plane.

Features

• Open a dataset root and browse frames from every camera at once
• Draw and edit boxes per camera; link objects with global IDs across views
• BEV panel for top-down placement and ground-point editing
• Annotations stored under annotations_positions/ as JSON (one file per frame)
• Optional preprocessing (YOLO-based detection + cross-camera matching) to bootstrap labels

Dataset layout

In the app, use File → Open Dir and select the root folder of your dataset. The expected layout is:

dataset_root/
├── Image_subsets/           # frame images, grouped by camera
│   ├── 1/                   # camera folder (name can be 1, 2, … or similar)
│   │   ├── 00000.jpg
│   │   └── ...
│   ├── 2/
│   └── ...
├── calibrations/            # one folder per camera
│   ├── Camera1/
│   ├── Camera2/
│   └── ...
└── annotations_positions/   # written when you save (may not exist yet)
    ├── 00000.json
    └── ...

• Image_subsets: each subfolder is one camera; image filenames should sort consistently by frame index.
• calibrations: camera intrinsics/extrinsics used for BEV projection (folder names like Camera1, Camera2, …).
• annotations_positions: global per-frame annotations; created and updated by the tool.

If both Image_subsets/ and calibrations/ are present, ArgusTrack switches to multi-camera mode automatically.

Start

Requirements: Python 3.9+ and uv.

git clone https://github.com/UARK-AICV/argustrack.git
cd argustrack
uv sync
uv run argustrack

Then File → Open Dir → choose dataset_root.

User settings are saved to ~/.argustrackrc on first run. Logs:

• Windows: %LOCALAPPDATA%\argustrack\argustrack.log
• Linux / macOS: ~/.cache/argustrack/argustrack.log

Optional: detection-assisted preprocessing

For automatic box proposals before manual refinement:

uv pip install ultralytics opencv-python

Use Preprocess from the menu after opening a dataset (requires the layout above).

Acknowledgement

This tool is built from wkentaro/labelme.