CNNGeometric PyTorch implementation
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
data added resnet model Nov 29, 2017
geotnf
image Initial commit Sep 4, 2017
model added resnet model Nov 29, 2017
training_data removed files from gitignore Sep 4, 2017
util
.gitignore moved trained models to separate server Sep 21, 2017
README.md
demo.py added resnet model Nov 29, 2017
demo_notebook.ipynb added resnet model Nov 29, 2017
eval_pf.py added resnet model Nov 29, 2017
train.py

README.md

CNNGeometric PyTorch implementation

This is the implementation of the paper:

I. Rocco, R. Arandjelović and J. Sivic. Convolutional neural network architecture for geometric matching. CVPR 2017 [website][arXiv]

using PyTorch (for MatConvNet implementation click here).

If you use this code in your project, please cite use using:

@InProceedings{Rocco17,
  author       = "Rocco, I. and Arandjelovi\'c, R. and Sivic, J.",
  title        = "Convolutional neural network architecture for geometric matching",
  booktitle    = "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition",
  year         = "2017",
}

Dependencies

  • Python 3
  • pytorch > 0.2.0, torchvision
  • numpy, skimage (included in conda)

Getting started

  • demo.py demonstrates the results on the ProposalFlow dataset
  • train.py is the main training script
  • eval_pf.py evaluates on the ProposalFlow dataset

Trained models

Using Streetview-synth dataset + VGG

  • [Affine], [TPS]
  • Results on PF: PCK affine: 0.472, PCK tps: 0.513, PCK affine+tps: 0.572

Using Pascal-synth dataset + VGG

  • [Affine], [TPS]
  • Results on PF: PCK affine: 0.478, PCK tps: 0.428, PCK affine+tps: 0.568

Using Pascal-synth dataset + ResNet-101

  • [Affine], [TPS]
  • Results on PF: PCK affine: 0.559, PCK tps: 0.582, PCK affine+tps: 0.676