Skip to content
Contextual Loss (CX) and Contextual Bilateral Loss (CoBi).
Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
contextual_loss
doc 🐛 badge fix? Oct 2, 2019
tests 🚀 add contents Sep 24, 2019
.gitignore
LICENSE
Pipfile 🆙 update documents Oct 2, 2019
Pipfile.lock 🆙 update documents Oct 2, 2019
README.md
setup.py 🚀 add contents Sep 24, 2019

README.md

Contextual Loss

PyTorch implementation of Contextual Loss (CX) and Contextual Bilateral Loss (CoBi).

Introduction

There are many image transformation tasks whose spatially aligned data is hard to capture in the wild. Pixel-to-pixel or global loss functions can NOT be directly applied such unaligned data. CX is a loss function to defeat the problem. The key idea of CX is interpreting images as sets of feature points that don't have spatial coordinates. If you want to know more about CX, please refer the original paper, repo and examples in ./doc directory.

Requirements

  • Python3.7+
  • torch & torchvision

Installation

pip install git+https://github.com/S-aiueo32/contextual_loss_pytorch.git

Usage

You can use it like PyTorch APIs.

import torch

import contextual_loss as cl
import contextual_loss.fuctional as F


# input features
img1 = torch.rand(1, 3, 96, 96)
img2 = torch.rand(1, 3, 96, 96)

# contextual loss
criterion = cl.ContextualLoss()
loss = criterion(img1, img2)

# functional call
loss = F.contextual_loss(img1, img2, band_width=0.1, loss_type='cosine')

# comparing with VGG features
# if `use_vgg` is set, VGG model will be created inside of the criterion
criterion = cl.ContextualLoss(use_vgg=True, vgg_layer='relu5_4')
loss = criterion(img1, img2)

Reference

Papers

  1. Mechrez, Roey, Itamar Talmi, and Lihi Zelnik-Manor. "The contextual loss for image transformation with non-aligned data." Proceedings of the European Conference on Computer Vision (ECCV). 2018.
  2. Mechrez, Roey, et al. "Maintaining natural image statistics with the contextual loss." Asian Conference on Computer Vision. Springer, Cham, 2018.

Implementations

Thanks to the owners of the following awesome implementations.

You can’t perform that action at this time.