On the Design of Privacy-Aware Cameras: A Study on Deep Neural Networks

This repository contains Pytorch training and evaluation code for semantic segmentation using DeepLabv3. It was partially inspired from aerial_mtl.

For details, see On the Design of Privacy-Aware Cameras: a Study on Deep Neural Networks.

This paper was accepted on ECCV 2022 IWDSC (International Workshop on Distributed Cameras).

If you use this code for a paper, please cite:

@article{carvalho2022privacy,
  title={On the Design of Privacy-Aware Cameras: a Study on Deep Neural Networks},
  author={Marcela Carvalho and Oussama Ennaffi and Sylvain Chateau and Samy Ait Bachir},
  journal={available-soon},
  year={2022}
}

Code will be available soon.

Usage

Clone this repository locally:

git clone https://github.com/upciti/privacy-by-design-semseg

We use poetry to... so, to install all the dependencies, we suggest using python ... and pyenv here is the explanation.

Compared to conda, which is the togo solution in the field, poetry has the advantage of... However there are some problems when we use some library that is not on their repo.

poetry install  # reads the pyptoject.toml file, resolves dependencies and installs them
poetry shell  # to activate the virtual environment in the current shell
poe poe-torch-cuda11  # to install Pytorch with CUDA 11.6

Data preparation

To be added.

Model zoo

name	url
infocus_color	download_weight
infocus_gray	download_weight
defocus_color	download_weight
defocus_gray	download_weight

We suggest to save weights inside a folder called checkpoints/model_name.

For the command lines we mention in this README, we consider you saved infocus_color weights in ./checkpoints/infocus_color/0300.pth.tar.

Visualisation

We adopt visdom to visualise training artifacts and metrics results. To run visdom, in another shell inside the project folder, run:

poetry shell
visdom -p $DISPLAY_PORT

Training

To train a pretrained DeepLabV3 with one of our weights with a single GPU, run:

python train_test_semseg.py --train --name name-of-the-new-project-with-pretrained-weights --dataroot PATH_TO_CITYSCAPES --batchSize $batch_size --nEpochs $end --display_id $display_id --port $port --use_resize --data_augmentation f f --resume

To train from scratch, run:

python train_test_semseg.py --train --name name-of-the-new-project --dataroot PATH_TO_CITYSCAPES --batchSize $batch_size --nEpochs $end --display_id $display_id --port $port --use_resize --data_augmentation f f 

Evaluation

To train a pretrained DeepLabV3 with one of our weights with a single GPU, run:

Generate images only:

python train_test_semseg.py --test --test_only --name infocus_color --save_samples --use_resize --display_id 0 --dataroot PATH_TO_DATA

Generate images and evaluation:

python train_test_semseg.py --test --test_metrics --name infocus_color --use_resize --display_id 0 --dataroot ./datasets/public_datasets/Cityscapes

As a result from this last run, you should get the outputs under the results file and the following metric results:

OA:  0.9387813619099232
AA:  0.7375342230955388
mIOU 0.6486886632431875

Generate only evaluation (resulting segmentation must be in the corresponding folder):

python train_test_semseg.py --test --evaluate_only --name infocus_color --use_resize --display_id 0 --dataroot ./datasets/public_datasets/Cityscapes

License

Code (scripts) are released under the MIT license.