locus-tag 0.1.2

A high-performance fiducial marker detector for robotics.

Locus (locus-vision)

The Project Goal

Locus is a high-performance fiducial marker detector (AprilTag & ArUco) written in Rust with zero-copy Python bindings. It targets a balance of low latency, high recall, and precise pose estimation.

[!WARNING] Experimental Status: Locus is an experiment on AI-assisted coding and the code is still actively iterated on. The API is subject to breaking changes. While performance is high on current benchmarks, it is not recommended for production systems yet. Photo-realistic benchmarks are in development under render-tag.

Performance (ICRA 2020 Dataset)

Detector	Recall	RMSE	Latency (avg)
Locus (Soft)	95.42%	0.26 px	129.1 ms
Locus (Hard)	83.90%	0.25 px	97.9 ms
AprilTag 3	62.34%	0.22 px	121.0 ms
OpenCV	33.16%	0.92 px	113.0 ms

Note: Locus' higher recall (detecting more challenging tags) correlates with its aggregate RMSE. On identical detections, Locus' precision is within +0.0024 px of AprilTag.

Quick Start

Install

pip install locus-tag
# Or from source
uv run maturin develop -r

Basic Usage

The simplest way to detect tags using default settings:

import cv2
import locus

# Load image in grayscale
img = cv2.imread("image.jpg", cv2.IMREAD_GRAYSCALE)

# Detect tags (defaults to AprilTag 36h11)
tags = locus.detect_tags(img)

for t in tags:
    print(f"ID: {t.id}, Center: {t.center}")

Advanced Features

For performance tuning, pose estimation, and specialized profiles, see the User Guide.

• Decimation: Speed up processing on high-res images.
• Soft-Decoding: Maximum recall for blurry or small tags (+11.5% boost).
• Pose Estimation: Advanced IPPE-Square and Probabilistic refinement.
• Profiles: Optimized settings for checkerboards and dense patterns.

Development & Benchmarking

Locus includes a rigorous suite to ensure detection quality and latency targets.

# Prepare datasets
uv run python scripts/locus_bench.py prepare

# Run evaluation suite
uv run python scripts/locus_bench.py run real --compare

Detailed instructions for profiling and regression testing are available in the Benchmarking Guide.

License

Dual-licensed under Apache 2.0 or MIT.