Code and its usage for reproducing our work FVRA-FDT

paper: Huang, J., S. Bhattacharjee, S., Marcel, S. and Kang, W.  Study of Full-View Finger Vein Biometrics on Redundancy Analysis and Dynamic Feature Extraction, IEEE Transactions on Information Forensics and Security

1. Architecture of FDT model

Architecture of the FDT model: (a) Overall architecture of FDT; (b) Dynamic Transformer module in FDT.

2. Structure of the code

3. Environment install

• Step 1: cd to the path of fdt.yaml file
• Step 2: use the following command for installing the environment

conda env create -f fdt.yaml

4. FVAB baseline (following the MFFV dataset paper)

Note: Here we just conduct the baseline for LFMB-3DFB dataset; The results of MFFV-N dataset are from the MFFV dataset paper

• Step 1: activate the fdt enviroment (or other bob environment) and cd to path of FVAB
• Step 2: use the following command for feature extraction

python feature_extraction.py -db /path/of/the/MFFV/dataset/ -fe /path/of/the/feature/you/want/to/save/

• Step 3: use the following command for single-view matching

python fvia.py -db /path/of/the/MFFV/dataset/ -fe /path/of/the/feature/you/had/saved/ -pt /path/of/the/protocols(PT folder)/ -rs RS/

• Step 4: use the following command for full-view fusion

python fusion_fvia.py -rs RS/ -fs RS/ # recommendation: both parameters may point to the same folder (i.e., output folder may be the same as the input folder).

5. FVRA experiments

• Step1: activate the fdt environment (or other bob environment) and cd to path of FVRA
• Step2: use the following command for single-view and full-view redundancy analysis of one example

python information_distribution.py -db /path/of/the/MFFV/dataset/ -rs RS/

• Step3: use the following command for general statistically redundancy analysis (all the first sample of the training set)

python general_statistic.py -db /path/of/the/MFFV/dataset/ -pt /path/of/the/PT/ -rs RS/

6. Training the FDT, Ablated FDT, comparison models on MFFV-N dataset

• Step 1: activate the fdt environment and cd to the path of train.py
• Step 2: use the following command for training the FDT

python train.py -m fdt -d MFFV-N -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/fdt_mffv/ 

• Step 3: use the following commands for training the Ablated FDT models

python train.py -m fdt_wo_ba -d MFFV-N -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/fdt_wo_ba_mffv/ # without BatchAtten
python train.py -m fdt_wo_lf -d MFFV-N -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/fdt_wo_lf_mffv/ # without LFFN
python train.py -m fdt_wo_mp -d MFFV-N -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/fdt_wo_mp_mffv/ # without MLP-P
python train.py -m fdt_wo_pe -d MFFV-N -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/fdt_wo_pe_mffv/ # without position emebdding
python train.py -m fdt_wo_dr -d MFFV-N -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/fdt_wo_dr_mffv/ # without dynamic reduce

• Step 4: use the following commands for training the comparison models

python train.py -m mvcnn_imp -d MFFV-N -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/mvcnn_ori_mffv/ # improved MVCNN 
python train.py -m mvt_imp -d MFFV-N -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/mvcnn_ori_mffv/ # improved MVT

7. Training the FDT on LFMB-3DFB dataset

• Step 1: activate the fdt environment
• Step 2: use the following command for training the FDT on LFMB-3DFB dataset

python train.py -m fdt -d LFMB-3DFB -db /path/to/the/dataset/LFMB-3DFB/LFMB-3DFB_Pictures_Seged_Rectified/ -pt /path/to/the/PT/ -ep 5000 -rs RS/trained_models/fdt_lfmb3dfb/

8. Obtain the experimental results of all models and save them as ‘.csv’ files

• Step1: use the following command, then you will see all the results of comparison models are saved in RS/SNM/ folder.

python snm.py -m ['fdt_mffv', 'mvcnn_ori_mffv_lgpu', 'mvcnn_imp_mffv', 'mvt_ori_mffv', 'mvt_imp_mffv'] -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -rs RS/SNM/ 

• Step2: use the following command, the you will see all the results of ablated models are saved in RS/SNM/ folder.

python snm.py -m ['fdt_wo_pe_mffv', 'fdt_wo_mp_mffv', 'fdt_wo_lf_mffv', 'fdt_wo_dr_mffv', 'fdt_wo_ba_mffv'] -db /path/to/the/dataset/MFFV-N/ -pt /path/to/the/PT/ -rs RS/SNM/ 

9. Obtain the experiments results of complexity and save them as ‘.txt’ files

• Just use the following command, then you will see all the results of the complexity experiment are saved in RS/params_flops/ folder.

python pnf.py -rs RS/params_flops/

10. Trained models

• Since all the experiments are conducted on idiap's computer, I did not re-run them when writing this readme.
• All the trained models can be download via Baidu cloud disk: https://pan.baidu.com/s/1rrGUjeoL5QheTPCkt9vmuQ?pwd=tifs

11. Contact and collaboration

If you find this work helpful to you and want to collaborate with me, please contact me (Junduan Huang) via Email: runrunjun@163.com.

12. Warning: Note About the License:

All files made available in this repository is released under the BSD.3 license, which is described in the file LICENSE.txt .

13. Dataset

1. MFFV-N dataset:

• Paper: https://ieeexplore.ieee.org/document/10741574
• Repository: https://github.com/SCUT-BIP-Lab/MFFV

2. LFMB-3DFB dataset:

• Paper: https://ieeexplore.ieee.org/document/9484369, and https://ieeexplore.ieee.org/document/9795663/
• Repository: https://github.com/SCUT-BIP-Lab/LFMB-3DFB

14. Citation

@ARTICLE{11236466, author={Huang, Junduan and Bhattacharjee, Sushil and Marcel, Sébastien and Kang, Wenxiong}, journal={IEEE Transactions on Information Forensics and Security}, title={Study of Full-View Finger Vein Biometrics on Redundancy Analysis and Dynamic Feature Extraction}, year={2025(Early Access)}, keywords={Biometrics;Redundancy;Feature extraction;Fingers;Transformers;Imaging;Authentication;Taxonomy;Convolutional neural networks;Cameras;Biometrics;vein recognition;finger vein;full-view;dynamic Transformer}, doi={10.1109/TIFS.2025.3630891}}