jasmcaus/caer

Caer – A lightweight, GPU‑accelerated computer‑vision toolbox

What it is – Caer is a pure‑Python library that wraps fast GPU kernels to give you the most common image‑processing and video‑handling operations (resize, colour conversion, augmentations, etc.) without the heavy boilerplate of OpenCV or custom CUDA code. It is positioned as a research‑oriented alternative that is type‑checked, well‑documented and easy to extend.

Key components

Module	Purpose
caer	Core high‑performance functions (e.g., resize, rotate).
caer.color	Colour‑space conversions and utilities.
caer.data	Bundles a few standard test images (e.g., sunrise).
caer.path	Cross‑platform path helpers.
caer.preprocessing	Common pre‑processing steps for images.
caer.transforms	Augmentation pipelines (flip, crop, etc.).
caer.video	Simple video‑reading/writing utilities.

Why you might use it

• Want GPU speed for classic CV ops but don’t want to write CUDA yourself.
• Need a clean, type‑checked API that plays nicely with PyTorch/TensorFlow pipelines.
• Looking for a small dependency footprint compared with full OpenCV builds.

Installation

pip install --upgrade caer

The project supports Python 3.6+ and provides wheels for the most common platforms. Detailed build instructions are in the INSTALL.md file for users who need to compile from source.

Minimal example

import caer

# Load a bundled 640×427 test image
sunrise = caer.data.sunrise(rgb=True)

# Resize to 400×400 while keeping the aspect ratio
resized = caer.resize(
    sunrise,
    target_size=(400, 400),
    preserve_aspect_ratio=True,
)

The README shows the resulting image and points to a folder of more demos.

Documentation & community

• Full API docs are hosted on ReadTheDocs (caer.rtfd.io).
• Issue tracking, discussions, and contribution guidelines are provided via GitHub.
• The package is MIT‑licensed, so it can be used in both open‑source and commercial projects.

Who it’s for Students, hobbyists, and researchers who need a fast, Pythonic way to prototype vision pipelines without pulling in a heavyweight C++ library.

---

All information above is taken directly from the repository’s README.