ignacio-rocco / cnngeometric_pytorch

CNNGeometric PyTorch implementation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

CNNGeometric PyTorch implementation

This is the implementation of the paper:

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

Dependencies

See requirements.txt

Demo

Please see the demo.py script or the demo_notebook.ipynb Jupyter Notebook.

Training

You can train the model using the train.py script in the following way:

python train.py  --geometric-model affine

For a full set of options, run python train.py -h.

Logging Configuration
  • For now it is implemented to log on TensorBoard just scalars of train and val loss
  • It is possible to specify a --logdir as a parameter, otherwise the logging folder will be named as the checkpoint one with _tb_logs as suffix
  • N.B. If is intended to use as logdir a GCP bucket it is necessary to install Tensorflow

Evaluation

You can evaluate the trained models using the eval.py script in the following way:

python eval.py  --model-1 trained_models/best_streetview_checkpoint_adam_hom_grid_loss_PAMI.pth.tar --eval-dataset pf

You can also evaluate a two-stage model in the following way:

python eval.py --model-1 trained_models/best_streetview_checkpoint_adam_hom_grid_loss_PAMI.pth.tar --model-2 trained_models/best_streetview_checkpoint_adam_tps_grid_loss_PAMI.pth.tar --eval-dataset pf

The eval.py scripts implements the evaluation on the PF-Willow/PF-PASCAL/Caltech-101 and TSS datasets. For a full set of options, run python eval.py -h.

Trained models

Model PF-Willow (PCK)
[Affine - VGG - StreetView] 48.4
[Homography - VGG - StreetView] 48.6
[TPS - VGG - StreetView] 53.8

BibTeX

If you use this code in your project, please cite us 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},
}

or

@Article{Rocco18,
  author = {Rocco, I. and Arandjelovi\'c, R. and Sivic, J.},
  title  = {Convolutional neural network architecture for geometric matching},
  journal = {{IEEE Transactions on Pattern Analysis and Machine Intelligence}},
  number = {41},
  pages = {2553--2567},
  year = {2018},
}

About

CNNGeometric PyTorch implementation

License:MIT License


Languages

Language:Python 50.9%Language:Jupyter Notebook 49.1%