anycv

One-liner computer vision — detect, classify, and segment in 3 lines of code.

anycv is a dead-simple computer vision inference library. It wraps the most popular pretrained models (YOLOv8 for detection, MobileNetV2/ResNet50 for classification, DeepLabV3 for segmentation) behind a one-liner API. Models are executed via ONNX Runtime for fast CPU inference (or GPU if available), auto-downloaded from Hugging Face Hub on first use, and cached locally. No PyTorch required at inference time.

Built by Viet-Anh Nguyen at NRL.ai.

Why anycv?

• One-liner API — anycv.detect("photo.jpg") returns bounding boxes immediately
• Plugin architecture — Register custom backends (TensorRT, OpenVINO, VLMs) via @register_backend
• Local-first — Models cached in ~/.cache/anycv/, zero network after first run
• Minimal core deps — Only onnxruntime, pillow, numpy; PyTorch is optional
• Production-ready — Type hints, dataclass results, streaming, batch inference

Installation

pip install anycv

For optional features:

pip install anycv[gpu]       # onnxruntime-gpu for CUDA inference
pip install anycv[torch]     # use PyTorch models directly
pip install anycv[vlm]       # Vision-LLM backend via anyllm (GPT-4o, Claude)
pip install anycv[all]       # everything

Python 3.8+ supported (tested on 3.8, 3.9, 3.10, 3.11, 3.12, 3.13)

Quick Start

import anycv

# 1. Object detection (YOLOv8n via ONNX Runtime — auto-downloaded from HF Hub)
detections = anycv.detect("street.jpg", model="yolov8n")
for d in detections:
    print(d.label, d.confidence, d.bbox)  # e.g. "person" 0.92 (x1,y1,x2,y2)

# 2. Image classification (MobileNetV2 ImageNet-1k via ONNX Runtime)
result = anycv.classify("cat.jpg", model="mobilenetv2")
print(result.top_k(5))                   # [(label, prob), ...]

# 3. Semantic segmentation (DeepLabV3 Pascal-VOC via ONNX Runtime)
mask = anycv.segment("scene.jpg", model="deeplabv3")
mask.save("mask.png")                    # per-pixel class map

Models & Methods

All models are distributed as ONNX files, auto-downloaded from Hugging Face Hub on first use and cached in ~/.cache/anycv/. Inference runs on ONNX Runtime (CPU by default; GPU via onnxruntime-gpu).

Object Detection

Exported from Ultralytics YOLOv8.

Image Classification

Semantic Segmentation

Vision-LLM Backend (optional)

When installed with anycv[vlm], you can use multi-modal LLMs for detection/classification via natural-language prompts:

# Uses anyllm to call GPT-4o, Claude 3.5 Sonnet, Gemini, or local LLaVA
result = anycv.classify("x-ray.jpg", backend="vlm", model="gpt-4o",
                        prompt="Classify findings: normal, pneumonia, or other")

Preprocessing pipeline

Every backend applies: letterbox resize -> BGR/RGB normalization -> mean/std standardization -> NCHW transpose. All handled automatically.

API Reference

Result dataclasses

@dataclass
class Detection:
    label: str
    confidence: float
    bbox: tuple[float, float, float, float]  # x1, y1, x2, y2
    class_id: int

CLI Usage

anycv detect street.jpg --model yolov8n --conf 0.3 --save annotated.jpg
anycv classify cat.jpg --top 5
anycv segment scene.jpg --out mask.png
anycv list-models

Examples

Batch inference over a folder

import anycv
from pathlib import Path

# Load once, reuse across images (warm model, no reload)
model = anycv.load_model("yolov8s")
for img in Path("images/").glob("*.jpg"):
    dets = model.detect(img, conf=0.4)
    print(img.name, len(dets), "objects")

Use a custom confidence + visualization

import anycv

dets = anycv.detect("crowd.jpg", model="yolov8m", conf=0.5)
annotated = anycv.draw("crowd.jpg", dets)   # Pillow image with boxes
annotated.save("out.jpg")

Register a custom backend

from anycv import register_backend, BaseDetector

@register_backend("detect", "my_tensorrt")
class TensorRTDetector(BaseDetector):
    def predict(self, image): ...

anycv.detect("photo.jpg", model="my_tensorrt")

License

MIT (c) Viet-Anh Nguyen