spatialhub

A lightweight computer vision library built around ONNX Runtime.

spatialhub is a open-source computer vision library that collects useful perception models behind a simple and consistent Python API. The focus is on lightweight inference, minimal dependencies, and easy deployment.

Status: Work in progress. The library is actively being developed by a single maintainer. APIs may change, and only a small number of models are currently available.

Why this project?

Many computer vision models are easy to use in research but harder to deploy because of large dependencies and different APIs.

The goal of spatialhub is to provide a simpler alternative by:

• Using ONNX Runtime for inference.
• Avoiding a direct PyTorch dependency where possible.
• Downloading model weights automatically.
• Providing a consistent interface across different models.
• Supporting headless environments with opencv-python-headless.

The project is still growing, and the feature set will expand over time.

Available Models

Currently supported:

• EfficientLoFTR — Semi-dense local feature matching.
• DepthAnything-3 — Monocular relative and metric depth estimation.

Roadmap

Some models and utilities planned for future releases include:

• Monocular depth estimation
• Pose estimation
• Object tracking
• Point cloud utilities

These are ideas, not promises, and development depends on available time.

Installation

Requires Python 3.12+.

Install from PyPI:

pip install spatialhub

Optional GPU support:

pip install "spatialhub[gpu]"

Quick Start

All models follow a consistent, unified API pattern: Initialize the engine, pass inputs, and receive structured predictions.

1. Feature Matching (EfficientLoFTR)

from spatialhub import EfficientLoFTR

matcher = EfficientLoFTR()
result = matcher.match("image1.jpg", "image2.jpg", max_dim=1024,)

result.visualize(
    top_k=50,
    save_path="output.png",
)

2. Depth & Pose Estimation (Depth Anything 3)

import cv2
from spatialhub import DepthAnything3

estimator = DepthAnything3(model_name="da3_base")
depth_result = estimator.estimate_depth(images=["view1.png", "view2.png"])

# Render colored depth map
depth_viz = estimator.visualize(depth_result.depth[0])
cv2.imwrite("depth.png", depth_viz)

Project Goals

• Simple Python API
• Lightweight inference
• Consistent model interfaces
• Automatic model weight management
• Easy deployment in scripts, servers, and containers

License

The core spatialhub library is licensed under the Apache 2.0 License.

Model architectures and pretrained weights keep the licenses provided by their original authors. See the documentation for each model for licensing details.