Skip to content

This repository contains the reference implementation for our proposed Convolutional CRFs.

License

Notifications You must be signed in to change notification settings

MarvinTeichmann/ConvCRF

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

36 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ConvCRF

This repository contains the reference implementation for our proposed Convolutional CRFs in PyTorch (Tensorflow planned). The two main entry-points are demo.py and benchmark.py. Demo.py performs ConvCRF inference on a single input image while benchmark.py compares ConvCRF with FullCRF. Both scripts output plots similar to the one shown below.

Example Output

Requirements

Plattform: Linux, python3 >= 3.4 (or python2 >= 2.7), pytorch 0.4 (or pytorch 0.3 + pyinn), cuda, cudnn

Python Packages: numpy, imageio, cython, scikit-image, matplotlib

To install those python packages run pip install -r requirements.txt or pip install numpy imageio cython scikit-image matplotlib. I recommand using a python virtualenv.

Optional Packages: pyinn, pydensecrf

Pydensecrf is required to run FullCRF, which is only needed for the benchmark. To install pydensecrf, follow the instructions here or simply run pip install git+https://github.com/lucasb-eyer/pydensecrf.git. Warning Running pip install git+ downloads and installs external code from the internet.

PyINN allows us to write native cuda operations and compile them on-the-fly during runtime. PyINN is used for our initial ConvCRF implementation and required for PyTorch 0.3 users. PyTorch 0.4 introduces an Im2Col layer, making it possible to implement ConvCRFs entirely in PyTorch. PyINN can be used as alternative backend. Run pip install git+https://github.com/szagoruyko/pyinn.git@master to install PyINN.

Execute

Demo: Run python demo.py data/2007_001288_0img.png data/2007_001288_5labels.png to perform ConvCRF inference on a single image. Try python demo.py --help to see more options.

Benchmark: Run python benchmark.py data/2007_001288_0img.png data/2007_001288_5labels.png to compare the performance of ConvCRFs to FullCRFs. This script will also tell you how much faster ConvCRFs are. On my system ConvCRF7 is more then 40 and ConvCRF5 more then 60 times faster.

Citation

If you benefit from this project, please consider citing our paper.

TODO

  • Build a native PyTorch 0.4 implementation of ConvCRF
  • Provide python 2 implementation
  • Build a Tensorflow implementation of ConvCRF

About

This repository contains the reference implementation for our proposed Convolutional CRFs.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages