edgefirst-image

GPU-accelerated image preprocessing for edge AI — colour conversion, resize, letterbox, rotation, tiled inference and annotation drawing, in one call per frame.

Part of the EdgeFirst HAL

edgefirst-image is one of five Python packages built from the EdgeFirst Hardware Abstraction Layer.

The EdgeFirstAI/hal repository is the home for all of them — source, issue tracker, architecture documentation and release notes.

Installation

pip install edgefirst-image

Requires Python 3.8 or newer; edgefirst-tensor and NumPy are installed automatically. Wheels are published for Linux (x86_64, aarch64), macOS (arm64), and Windows (x86_64). This package does not depend on edgefirst-decoder.

Packages install under the PEP 420 edgefirst.* namespace, so the import is edgefirst.image.

A single OpenGL ES engine backs this package on every platform — Linux via native EGL and DMA-BUF, macOS and iOS via ANGLE over Metal and IOSurface, Android via native EGL and AHardwareBuffer, Windows via ANGLE over Direct3D 11 with PBO transfers — alongside NXP G2D on i.MX and a portable CPU fallback everywhere else. Backend selection is automatic; no code changes are needed to move between them. The Windows wheel bundles ANGLE's libEGL.dll and libGLESv2.dll next to the extension module, so no setup is required there; EDGEFIRST_ANGLE_ADAPTER picks the GPU on multi-adapter machines.

Quick start

Preprocess a frame into a model input tensor. convert() performs colour conversion, resize, letterboxing, rotation and cropping in a single GPU pass:

import numpy as np
from edgefirst.image import ImageProcessor, PixelFormat

processor = ImageProcessor()

# Allocate once, outside the loop.
src = processor.create_image(1280, 720, PixelFormat.Rgb, "uint8", "readwrite")
model_input = processor.create_image(640, 640, PixelFormat.Rgb, "uint8", "readwrite")

# Fill `src` from your capture source.
with src.map() as view:
    np.frombuffer(view, dtype=np.uint8)[:] = 200

# Omit letterbox= to stretch instead of preserving aspect ratio.
processor.convert(src, model_input, letterbox=[114, 114, 114, 255])

with model_input.map() as view:
    frame = np.frombuffer(view, dtype=np.uint8).reshape(640, 640, 3)

Use create_image() rather than allocating tensors yourself: it supplies DMA-BUF or PBO backing for zero-copy GPU import and GPU pitch alignment, which the plain allocator cannot guarantee.

PixelFormat and the other value types (TensorMemory, Region, the colour axis enums) are accepted from any edgefirst.* package, not just this one — they compare and hash equal across packages by value, so ==, dict keys and set membership all work regardless of which package's copy you pass. Tensors, Decoder and ProtoData cross packages too, through the capsule protocols. Importing PixelFormat from edgefirst.image when calling into this package is still good style for readability, not a requirement. See the Interoperability section below for the one thing that does not cross: isinstance against a concrete class.

What this package provides

Tiled inference (SAHI)

Small objects vanish when a high-resolution frame is squeezed into a 640×640 model input. The tiling APIs render an overlapping tile grid in a single GPU pass and merge per-tile detections back to full-frame coordinates, so small-object recall improves without a second inference pipeline.

Interoperability

convert(), draw_decoded_masks(), draw_proto_masks(), materialize_masks() and the other ImageProcessor entry points accept a tensor or ProtoData from any edgefirst.* package. Fused decode+draw lives on Decoder.draw_onto. Each extension module registers its own type objects (PyO3 issue #1444 — isinstance across packages is always False), so acceptance goes through duck-typed capsule protocols instead:

# CORRECT — works regardless of which edgefirst.* package produced obj
if hasattr(obj, "__edgefirst_tensor__"):
    ...

# WRONG — always False for a tensor from a sibling package
if isinstance(obj, edgefirst.codec.Tensor):
    ...

edgefirst.image.EdgeFirstTensorExportable, EdgeFirstDecoderExportable and EdgeFirstProtoDataExportable are typing.Protocol classes (re-exported here from edgefirst.tensor / edgefirst.decoder) you can annotate a cross-package parameter with. See crates/python-common/INTEROP.md for the full protocol — capsule names, lifetime and ownership rules, and versioning.

Versioning and changelog

All five edgefirst-* packages are versioned and released together with the HAL itself, so a given version number refers to the same source tree in every language. Because of that there is no per-package changelog: release notes for every version live in the single CHANGELOG.md in the hal repository, which follows Keep a Changelog and Semantic Versioning.

Links

• Changelog — release notes for all packages
• Source — the hal monorepo
• Issue tracker
• Package documentation — the underlying Rust crate
• EdgeFirst

License

Apache-2.0