This is the repository for our BMVC 2022 paper titled MorphPool: Efficient Non-linear Pooling & Unpooling in CNNs.
BMVC
arXiv

A requirements file is available to retrieve all dependencies. Create a new python environment and install using:

pip install -r requirements.txt

This paper makes available CUDA implementations for MorphPool, since morphological operations have not yet been implemented in CUDA for neural networks. The operations can be compiled using

cd morphology
sh setup.sh

This requires you to have installed all dependencies, and have the CUDA toolkit installed on your system. After compilation, some unit tests are run to check whether compilation was successful.

You can train any model using train.py by specifying your down and up-sampling operation. See options.py for the available operations.
In general, it should look like:

python train.py --pool_method param_pool --pool_ks 3 --unpool_method param_morph_unpool --unpool_ks 5

At the end of training, the weights are saved automatically to output. From these model files, you can test performance on the test sets using test.py. This is done by invoking test.py with the same arguments you used during training.
For example, the model that was trained using the command above, can be tested by:

python test.py --pool_method param_pool --pool_ks 3 --unpool_method param_morph_unpool --unpool_ks 5

This repository allows you to run our code on three semantic segmentation datasets: NYU, SUN-RGBD and 2D3D-S. Check out loaders to download and setup the datasets for your system.
After you have done so, you can train a model on a specific dataset using the --dataset argument for [ nyu | sun | 2d3ds ].
For example

python train.py --dataset sun

Note that you have to change the value of DATA_FOLDER on line 2 in config.py to the path to the folder you keep your datasets.

Finally, it is good to note that we use the simple network UpDownNet as introduced in the paper. It was meant to isolate changes in performance due to sampling, not reach the SOTA in semantic segmentation. Consequently, it has relatively few parameters, is not pre-trained on ImageNet, has no specific semantic modules, and does not have to be trained for many iterations. Feel free to apply the (morphological) up and down-sampling in more complex architectures.

If this work was useful for your research, please consider citing:

@inproceedings{groenendijk2022morphpool,
author    = {Rick Groenendijk and Leo Dorst and Theo Gevers},
title     = {MorphPool: Efficient Non-linear Pooling & Unpooling in CNNs},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year      = {2022},
url       = {https://bmvc2022.mpi-inf.mpg.de/0056.pdf}
}