corrmatch 0.2.0

CPU-first template matching library for grayscale images

corrmatch-py

Python bindings for corrmatch - CPU-first template matching library.

Installation

From source (development)

# Install maturin
pip install maturin

# Build and install in development mode
cd corrmatch-py
maturin develop

# Or build a release wheel
maturin build --release
pip install target/wheels/corrmatch-*.whl

Usage

Quick one-shot matching

import numpy as np
import corrmatch

# Load your images as numpy arrays (2D uint8)
image = np.array(...)      # shape: (height, width)
template = np.array(...)   # shape: (height, width)

# Simple matching (translation only)
result = corrmatch.match_template(image, template)
print(f"Found at ({result.x}, {result.y}) with score {result.score}")

# With rotation
result = corrmatch.match_template(image, template, rotation="enabled")
print(f"Found at ({result.x}, {result.y}), angle={result.angle_deg}deg, score={result.score}")

Efficient repeated matching

For matching the same template against multiple images:

import corrmatch

# Create and compile template once
tpl = corrmatch.Template(template_array)
compiled = tpl.compile()  # With rotation support
# Or: compiled = tpl.compile_no_rotation()  # Faster, no rotation

# Create matcher
matcher = compiled.matcher()

# Match against multiple images
for image in images:
    result = matcher.match_image(image)
    # Or: results = matcher.match_topk(image, k=5)  # Top 5 matches

Configuration

# Compile config (rotation parameters)
compile_cfg = corrmatch.CompileConfig(
    max_levels=6,           # Pyramid levels
    coarse_step_deg=10.0,   # Initial angle step
    min_step_deg=0.5,       # Finest angle step
    fill_value=0,           # Fill for rotated edges
)
compiled = tpl.compile(compile_cfg)

# Match config
match_cfg = corrmatch.MatchConfig(
    metric="zncc",          # "zncc" or "ssd"
    rotation="enabled",     # "enabled" or "disabled"
    parallel=True,          # Use rayon parallelism
    beam_width=8,           # Candidates per level
    nms_radius=6,           # Non-maximum suppression radius
)
matcher = compiled.matcher(match_cfg)

Loading from files

# Load template from image file
tpl = corrmatch.Template.from_file("template.png")

Visualization (matplotlib)

CorrMatch includes a small visualization helper to inspect results in an interactive matplotlib window (rotated detection frame + template + matched patch + deskew + diff).

Install extra dependencies:

pip install -e ".[viz]"

Run from Python:

import corrmatch
import corrmatch.viz as viz

match_cfg = corrmatch.MatchConfig(rotation="enabled", metric="zncc")
fig, matches = viz.match_and_show(image, template, topk=5, match_cfg=match_cfg)

Or as a CLI:

corrmatch-viz --image image.png --template template.png --rotation enabled --topk 5

Angle convention: positive angles are clockwise (x right, y down).

API Reference

Classes

• Template: Grayscale template image
• CompiledTemplate: Pre-compiled template with pyramids
• Matcher: Template matcher for coarse-to-fine search
• Match: Match result with x, y, angle_deg, score
• CompileConfig: Template compilation settings
• MatchConfig: Matching parameters

Functions

• match_template(image, template, ...): One-shot matching

Building

Requires:

• Rust toolchain (1.88+)
• Python 3.11+
• maturin (pip install maturin)

cd corrmatch-py
PYO3_USE_ABI3_FORWARD_COMPATIBILITY=1 maturin develop