Official implementation of the denoising (PyTorch) and correspondence classification (Tensorflow) N3NET, that will be published in our NeurIPS paper:

Tobias Plötz and Stefan Roth, Neural Nearest Neighbors Networks, Advances in Neural Information Processing Systems (NeurIPS), 2018

Preprint: https://arxiv.org/abs/1810.12575

NeurIPS Proceedings: https://papers.nips.cc/paper/7386-neural-nearest-neighbors-networks

Contact: Tobias Plötz ([email protected])

The denoising code is tested with Python 3.6, PyTorch 0.4.1 and Cuda 8.0 but is likely to run with newer versions of PyTorch and Cuda.

To install PyInn run

```
pip install git+https://github.com/szagoruyko/pyinn.git@master
```

Further requirements can be installed with

```
pip install -r requirements.txt
```

Update March 28, 2019 Since tensor comprehensions is not maintained anymore, we provide a memory and time efficient implementation of an indexed matrix multiplication. To build the corresponding cuda kernel please cd into lib and run

```
python setup.py install
```

Please download the BSDS500, Urban100 and Set12 datasets by cd'ing into datasets/ and using the scripts provided therein. If you want to train your own Poisson-Gaussian denoising model, please additionally download the DIV2k dataset and the Waterloo dataset.

For setting up the correspondence classification code, please clone the repository (https://github.com/vcg-uvic/learned-correspondence-release), follow their instructions to setup your environment and copy the files located in src_correspondence/. This can also conveniently be done using the script src_correspondence/clone_CNNet.sh.

To run the following commands, please cd into src_denoising.

To train a new model run:

```
python main.py <options>
```

To test your model run:

```
python main.py --eval --eval_epoch <epoch> --evaldir <dir>
```

To test our pretrained networks run:

```
python main.py --eval --eval_epoch 51 --evaldir pretrained_sigma<25|50|70>
```

A note on pretrained models. The pretrained models will give slightly different results than in the paper on Set5 due to differences in the random seed. On the other, larger datasets, results are as in the paper. Furthermore, in the paper we used a strong learning rate decay. Training our model again with a slower decay yields better results and we will add new pretrained models soon.

To run the following commands, please cd into src_correspondence/CNNet.

To train a new model run:

```
python main.py --run_mode=train --net_arch=nips_2018_nl <options>
```

We provide two pretrained models. One is trained on the Brown indoor dataset, the other is trained on the St. Peters outdoor dataset. To test our pretrained models run:

```
python main.py --run_mode=test --net_arch=nips_2018_nl  --data_te=<"brown_bm_3_05"|"st_peters"|"reichstag"> --data_va=<"brown_bm_3_05"|"st_peters"|"reichstag">
--log_dir="<pretrained_brown|pretrained_stpeters>"
--test_log_dir=<dir>
```

The core of the PyTorch implementation is located in src_denoising/models/non_local.py which provides classes for neural nearest neighbors selection (NeuralNearestNeighbors), a domain agnostic N3Block (N3AggregationBase) and a N3Block tailored towards image data (N3Aggregation2D). The file src_denoising/models/n3net.py contains the N3Net module that uses the 2D N3Block as non-local processing layer.

The core of the Tensorflow implementation is located in src_correspondence/non_local.py which provides analogous functionality as above.

@inproceedings{Ploetz:2018:NNN,
  title     = {Neural Nearest Neighbors Networks},
  author    = {Pl\"otz, Tobias and Roth, Stefan},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2018}
}