Joint Bilateral Upsampling (JBU) for Python using CUDA

This repository is an implementation of the Joint Bilateral Upsampling (JBU) using CUDA. It can be installed as a Python package and used in a Python script with PyTorch tensors.

Installation

Prerequisites

The NVIDIA CUDA Compiler (NVCC) must be installed in order to build the project. This can be done by installing the CUDA Toolkit via apt or in a Conda env using the following commands:

# Via apt
sudo apt install nvidia-cuda-toolkit

# Via conda
conda install nvidia::cuda-toolkit

Package installation

You can install this package in an environnement using pip by typing the following command in the main folder of the package:

pip install .

Usage

You can call the JBU in a Python script as following:

import torch # Should be placed before the import of jbu_cuda. jbu_cuda cannot work without the torch import.
import jbu_cuda

# Parameters of the JBU
sigma_s:float = 4
radius:int = 10
sigma_r:float = 0.2

result:Tensor = jbu_cuda.upsample(guidance:Tensor, low_res:Tensor, radius, sigma_s, sigma_r)

The provided tensors should contain float between 0.0 and 1.0, with shape (B,C,h,w) for the low resolution image and (B,1,H,W) for the guidance image.

A example script is provided in the example/ folder.