Arm YOLO26n Python APIs

This package provides two intentionally separate interfaces to the fixed YOLO26n runtime:

Yolo26nTensorSession is the low-level interface for callers that already own preprocessing and postprocessing. It accepts and returns FP32 NumPy arrays.
Yolo26Detector is the high-level image interface. It accepts uint8 NumPy images and returns class-labelled detections in source-image coordinates.

Both interfaces accept the runtime library and converted model directory as separate, explicit inputs. The package validates the model and tensor contract before creating the native model handle. It reads model.bin once, verifies those bytes against the manifest, and passes them to the native runtime.

Raw tensor execution

Use Yolo26nTensorSession when integrating with an existing tensor pipeline:

import numpy as np

from arm_yolo26n import Yolo26nTensorSession

with Yolo26nTensorSession(
    runtime="/path/to/libarm-yolo26n.so.0",
    model="/path/to/converted-model",
) as session:
    model_input = np.random.default_rng().random(session.input_shape, dtype=np.float32)
    raw_prediction = session.execute(model_input)

Inputs must be contiguous or convertible to contiguous FP32 NCHW arrays with the exact shape recorded by the model. Outputs are FP32 NumPy arrays with the model's validated YOLO26n output shape.

Image detection

Use Yolo26Detector when starting with an image:

from arm_yolo26n import DetectorConfig, Yolo26Detector

config = DetectorConfig(confidence=0.25, iou=0.45)
with Yolo26Detector(
    runtime="/path/to/libarm-yolo26n.so.0",
    model="/path/to/converted-model",
    config=config,
) as detector:
    result = detector.detect(frame_bgr, pixel_format="BGR")
    for detection in result.detections:
        print(detection.class_name, detection.confidence, detection.box)

The detector handles letterboxing, channel conversion, normalization, native execution, non-maximum suppression, and scaling boxes back to source-image coordinates. It accepts interleaved BGR, RGB, and four-channel BGRX arrays. The BGRX path lets Picamera2 XRGB8888 buffers be consumed without repacking them into a second three-channel image.

Sessions and detectors own native resources and are intentionally non-reentrant. Use them as context managers, and use independent processes when concurrent isolation is required.

See the prerecorded-video example for repeatable workload benchmarking and the Picamera2 example for live camera integration.