Python Usage¶

Quickstart¶

import cv2
from opendetect import Detector

detector = Detector(model="rfdetr-m")

image = cv2.imread("input.jpg")
detections = detector.predict(image, color="bgr")
annotated = detector.annotate(image, detections, color="bgr")

cv2.imwrite("output.jpg", annotated)

File Helpers¶

from opendetect import Detector

detector = Detector(model="yolox-s")
detector.infer_image_file("input.jpg", output_path="output.jpg")
detector.infer_video_file("input.mp4", output_path="output.mp4", max_frames=300)

Choosing Models¶

You can load by registry model ID (recommended) or by implementation name.

registry IDs: rfdetr-n, rfdetr-s, rfdetr-m, rfdetr-l, yolox-tiny, yolox-s, yolox-m, yolox-l, yolox-x, rtmdet-n-hand, rtmdet-s, rtmdet-m, rtmdet-l, bytetrack-s, bytetrack-m, bytetrack-l
implementation names: rfdetr, yolox, rtmdet, bytetrack (maps to project defaults)

Runtime Controls¶

from opendetect import Detector

detector = Detector(
    model="rfdetr-l",
    hardware_acceleration=True,
    tensor_rt=True,
    mixed_precision=True,
)

hardware_acceleration=False forces CPU execution provider.
tensor_rt=True enables TensorRT provider selection when available.
mixed_precision=True enables FP16/TF32-capable settings where supported.

Class Filtering¶

By class IDs:

detector.set_class_filter([0, 1, 2])

By class names:

ids = detector.resolve_class_ids_from_names(["person", "car"])
detector.set_class_filter(ids)

Class IDs are normalized to 0-based foreground IDs across model families.

Model Metadata¶

from opendetect import list_models

for spec in list_models():
    print(spec.model_id, spec.implementation, spec.input_size)