WACV 2022 XAI4B Workshop Supplementary material - Source Code

Interpretable Deep Learning-Based Forensic Iris Segmentation and Recognition - WACV 2022

Requirements

Docker version 20.10.7 or higher

Instructions

Get a copy

Clone this repository into your machine:

git clone https://...

Build the container

In the folder where you cloned the repository, run the following command to build the container with the environment to run TLPIM:

docker build --build-arg USER_ID=$UID -t tlpim-cpu .

Building the container for the first time can take a while.

Download the models

Download the trained models for Mask R-CNN and TLPIM from these addresses:

TLPIM

Mask R-CNN

Place the downloaded files on the models folder of the project.

Run the container

After building the container, you can run it with:

docker run --rm -ti tlpim-cpu

The remaining tasks described in this README should be executed inside the tlpim-cpu container.

Iris Segmentation

Activating the environment

To activate the detectron2 environment, run:

conda activate detectron2

This command is necessary before running any segmentation task.

Run segmentation

Command to run segmentation on the provided samples:

python segmentation/maskrcnn/iris/finetune.py \
--weights models/maskrcnn-pm.pth \
--impath samples/ \
--genmasks individual \
--output ./

Train Mask R-CNN

To train Mask R-CNN on iris images, use segmentation/maskrcnn/iris/train_maskrcnn.py.

Check possible arguments with python segmentation/maskrcnn/iris/train_maskrcnn.py --help.

Fine tune Mask R-CNN

To train Mask R-CNN on highlights and wrinkles, use segmentation/maskrcnn/iris/finetune.py.

Check possible arguments with python segmentation/maskrcnn/iris/finetune.py --help.

Iris Recognition

Activate the environment

To activate TLPIM environment, run inside the container:

conda activate tlpim

This command is necessary before running any recognition task.

Run TLPIM

To use TLPIM to perform iris comparison use a command like this:

python  recognition/tlpim/evaluation.py predict \
--outpath ./ \
--weights ./models/triplet-postmortem.h5 \
--image_folder ./samples/cropped/ \
--testfiles ./samples/filelist.txt

Train TLPIM

You can train your own network using a command like this:

python recognition/tlpim/triplet_train.py \
-o ./ \
-p test- \
--type resnet \
-tf ./samples \
-mf ./ \
--lr 0.00001 \
--solver adam \
--batch_size 24 \
-vp 20 \
--min_iterations 20000 \
--patience 20 \
--update_base_model

Please note that training on the samples provided with this code will result in error, because there are not enough samples for formation of triplets, or even train/test partitioning.

Use python recognition/tlpim/triplet_train.py --help to view available arguments.

Acknowledgements

Parts of this code were based on the code kindly provided by the authors of https://doi.org/10.1109/IJCB48548.2020.9304939

akuehlka/xai4b_tlpim