Compositional Convolutional Neural Networks: A Deep Architecture with Innate Robustness to Partial Occlusion [CVPR-2020]

Compositional Convolutional Neural Networks: A Deep Architecture with Innate Robustness to Partial Occlusion
Adam Kortylewski, Ju He, Qing Liu, Alan Yuille
CVPR 2020

Release Notes

This is a port of our original code from Tensorflow to PyTorch. The code is a lot faster and cleaner compared to the original code base. The results are a little different from the ones reported in the paper. In particular, the performance is a little lower for low occlusion and higher for stronger occlusion. On average the results are slightly better than reported in the paper.

For now, we provide pretrained models for CompositionalNets trained from the VGG-16 pool5 layer. Training CompositionalNets for other backbones and layers should be possible but has not been extensively tested so far.

Installation

The code uses Python 3.6 and it is tested on PyTorch GPU version 1.2, with CUDA-10.0 and cuDNN-7.5.

Setup CompNet Virtual Environment

virtualenv --no-site-packages <your_home_dir>/.virtualenvs/CompNet
source <your_home_dir>/.virtualenvs/CompNet/bin/activate

Clone the project and install requirements

git clone https://github.com/AdamKortylewski/CompositionalNets.git
cd CompositionalNets
pip install -r requirements.txt

Download models

Download pretrained CompNet weights from here and copy them inside the models folder.
The repositroy contains a few images for the demo script. If you want to evaluate on the full datasets used in our paper you need to download the data here and copy it inside the data folder.

Demo

CompNets require a tight crop of the object in the image. We provide sample images in the demo folder which are taken from MS-COCO.

Run the demo code

python Code/demo.py

Our demo script classifies the images from the demo folder, extracts the predicted location of occluders, and writes the results back into the demo folder.

Evaluate the classification performance of a model

Run the following command in the terminal to evaluate a model on the full test dataset:

python Code/test.py

Evaluate the occluder localization performance of a model

If you want to test occluder localization run:

python Code/eval_occlusion_localization.py

This will output qualitative occlusion localization results for each image and a quantitative analysis over all images as ROC curve.

Initializing CompositionalNet Parameters

We initialize CompositionalNets (i.e. the vMF kernels and mixture models) by clustering the training data. In particular, we initialize the vMF kernels by clustering the feature vectors:

python Initialization_Code/vMF_clustering.py

Furthermore, we initialize the mixture models by EM-type learning. The initial cluster assignment for the EM-type learning is computed based on the similarity of the vMF encodings of the training images. To compute the similarity matrices use:

python Initialization_Code/comptSimMat.py

As this process takes some time we provide precomputed similarity matrices here, you need to copy them into the 'models/init_vgg/' folder. Afterwards you can compute the initialization of the mixture models by executing:

python Initialization_Code/Learn_mix_model_vMF_view.py

Referencing CompositionalNets

Please cite the following papers if you use the code directly or indirectly in your research projects.

@inproceedings{CompNet:CVPR:2020,
  title = {Compositional Convolutional Neural Networks: A Deep Architecture with Innate Robustness to Partial Occlusion},
  author = {Kortylewski, Adam and He, Ju and Liu, Qing and and Yuille, Alan},
  booktitle = {Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
  month = jun,
  year = {2020},
  month_numeric = {6}
}

@article{kortylewski2021compositional,
  title={Compositional convolutional neural networks: A robust and interpretable model for object recognition under occlusion},
  author={Kortylewski, Adam and Liu, Qing and Wang, Angtian and Sun, Yihong and Yuille, Alan},
  journal={International Journal of Computer Vision},
  volume={129},
  number={3},
  pages={736--760},
  year={2021},
  publisher={Springer}
}

Relation to our Prior Work

With very small modifications this code would also enable the learning and testing of CompositionalNets with Bernoulli distributions as proposed in our previous work:

@inproceedings{Combining:WACV:2020,
title = {Combining Compositional Models and Deep Networks For Robust Object Classification under Occlusion},
author = {Kortylewski, Adam and Liu, Qing and Wang, Huiyu and Zhang, Zhishuai and Yuille, Alan},
booktitle = {IEEE Winter Conference on Applications of Computer Vision (WACV)},
month = mar,
year = {2020},
month_numeric = {3}
}

Contact

If you have any questions you can contact Adam Kortylewski.

Acknowledgement

We thank Zhishuai Zhang for helping us speed up and clean the code for the release.

AdamKortylewski / CompositionalNets