ovkit

ovkit

A simple Python inference API for OpenVINO — one import, one Model class, a callable object, and clean Results, with AUTO/NPU devices, async throughput, and INT8 quantization.

from ovkit import Model

model = Model("rtdetr_r50")            # name -> auto download / convert / cache
for r in model("img.jpg", conf=0.25):  # __call__ == predict
    print(r.boxes.xyxy, r.boxes.conf, r.boxes.cls)
    r.save("out.jpg")

🚀 Get started

Install, load a model, run prediction, pick a device.

Usage

📖 Cookbook

Copy-paste call examples for every feature.

Cookbook — calling every feature

🧩 API reference

Model, Results, registry, adapters, image ops.

API reference

Capabilities

A capability name gives you the answer, not the plumbing — one Model(...) call chains a detector with the models that describe what it found.

from ovkit import Model, list_pipelines

list_pipelines()                                   # every capability, described
Model("face_analyze")("group.jpg")[0].summary()    # '2 faces: age 31 · male 98% · happy 92%, ...'
Model("read_text")("sign.jpg")[0].text             # 'STOP AHEAD'
Model("track")(0)                                  # webcam, ids stable across frames

Model(...)	Answers	Chains
face_analyze	ages, genders and emotions of every face	face detection + age/gender + emotion
person_analyze	what each person wears or carries	person detection + attributes
vehicle_analyze	type and colour of each vehicle	vehicle detection + attributes
read_text	every word, in reading order	text detection + recognition
track	a stable id per object across frames	detection + IoU association
gaze	where a face is looking	detection + landmarks + head pose + gaze
face_match	who this is, from your own gallery	embedding + cosine matching
scene	one sentence about the whole picture	detection + segmentation + faces
read_plate	number plates, and the car each is on	plate detection + OCR + vehicle attributes
drowsiness	eyes shut too long, or a nodding head	face + landmarks + eye state + head pose, over time
gesture	hand gestures from motion	sign-language model over a rolling 8-frame clip
attention	which object a person is looking at	gaze + object detection
anonymize	the picture with faces (and plates) removed	detection + redaction

What it does

🎯 Vision tasks

Detection (DETR / SSD / YOLO), classification, segmentation (semantic + instance), pose, OCR — model(img) → typed Results.

⚙️ Any model

Generic raw-tensor fallback + low-level model.infer(tensors) for NLP / audio / time-series.

💬 GenAI

LLM / Whisper (STT) / TTS via ovkit.genai.pipeline(...) (openvino-genai).

📦 Auto everything

Auto download + IR convert + cache, task auto-detection, input auto-routing (image / .npy / .wav), INT8 quantization.

ovkit is Apache-2.0 and stays license-clean: it never bundles or downloads AGPL-licensed model stacks or non-commercial weights. See the license policy.