Introduction

pyvisim is a computer vision library for computing image similarities using traditional and deep learning methods.

Overview

The goal of pyvisim is to become the largest collection of image similarity metrics, varying from traditional methods like PSNR, SSIM, Fisher Vectors, and VLAD to deep learning methods like CLIP and Siamese Networks. Furthermore, advanced image similarity search and reranking algorithms are provided to allow users to refine the search results as desired.

Currently, one would need to install numerous libraries just to everything mentioned above (for example, scikit-image + opencv-python for Fisher Vectors, SSIM, open-clip for CLIP Embedder, and faiss for Approximate Nearest Neighbors Search). pyvisim attempts to close this gap by implementing as many metrics as possible using only numpy, scipy (for conventional metrics), and torch (for deep learning metrics) as core dependencies, plus making them more user-friendly with a simple Object-Oriented code design.

Installation

To install the slim version (without deep learning features):

pip install pyvisim

Additional features include (note: these pull in heavy dependencies like torch):

# For deep learning features and the OxfordFlowerDataset
pip install "pyvisim[nn]"

torch and numpy images supported

All Similarity Metrics in pyvisim accept images as either numpy.ndarray or torch.Tensor. The outputs are, however, always numpy.ndarray. Make sure the dimensions and data ranges are passed correctly. See This Document for more information.