trimodal_person_verification

This repository contains the code, and preprocessed dataset featured in "A Study of Multimodal Person Verification Using Audio-Visual-Thermal Data".

Person verification is the general task of verifying person’s identity using various biometric characteristics. We study an approach to multimodal person verification using audio, visual, and thermal modalities. In particular, we implemented unimodal, bimodal, and trimodal verification systems using the state-of-the-art deep learning architectures and compared their performance under clean and noisy conditions.

Dependencies

pip install -r requirements.txt

Dataset

In this work, we utilized the SpeakingFaces dataset to train, validate, and test the person verification systems. SpeakingFaces is a publicly available multimodal dataset comprised of audio, visual, and thermal data streams. The preprocessed data used for our experiments can be downloaded from Google Drive.

The data directory contains the compressed version of the preprocessed data used for the reported experiments. For each utterance, only the first frame (visual and thermal) is selected. The train set is split into 5 parts that should be extracted into the same location.

The data/metadata subdirectory contains lists prepared for the train, validation, and test sets following the format of VoxCeleb. In particular, the train list contains the paths to the recordings and the corresponding subject identifiers. The validation and test lists consist of randomly generated positive and negative pairs. For each subject, the same number of positive and negative pairs were selected. In total, the numbers of pairs in the validation and test sets are 38,000 and 46,200, respectively.

Note, to run noisy training and evaluation, you should first download the MUSAN dataset.

See trainSpeakerNet.py for details on where the data should be stored.

Training examples : clean data

Unimodal models

python trainSpeakerNet.py --model ResNetSE34Multi --modality wav --log_input True --trainfunc angleproto --max_epoch 1500 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.01 --seed 1 --save_path exps/wav/exp1

python trainSpeakerNet.py --model ResNetSE34Multi --modality rgb --log_input True --trainfunc angleproto --max_epoch 600 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.01 --seed 1 --save_path exps/rgb/exp1

python trainSpeakerNet.py --model ResNetSE34Multi --modality thr --log_input True --trainfunc angleproto --max_epoch 600 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.01 --seed 1 --save_path exps/thr/exp1

Multimodal models

python trainSpeakerNet.py --model ResNetSE34Multi --modality wavrgb --log_input True --trainfunc angleproto --max_epoch 600 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.01 --seed 1 --save_path exps/wavrgb/exp1

python trainSpeakerNet.py --model ResNetSE34Multi --modality wavrgbthr --log_input True --trainfunc angleproto --max_epoch 600 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.1 --seed 1 --save_path exps/wavrgb/exp1

Training examples : noisy data

Unimodal models

python trainSpeakerNet.py --model ResNetSE34Multi --modality wav --noisy_train True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --max_epoch 1500 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.001 --seed 1 --save_path exps/wav/exp2

python trainSpeakerNet.py --model ResNetSE34Multi --modality rgb --noisy_train True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --max_epoch 600 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.01 --seed 1 --save_path exps/rgb/exp2

python trainSpeakerNet.py --model ResNetSE34Multi --modality thr --noisy_train True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --max_epoch 600 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.01 --seed 1 --save_path exps/thr/exp2

Multimodal models

python trainSpeakerNet.py --model ResNetSE34Multi --modality wavrgb --noisy_train True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --max_epoch 600 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.01 --seed 1 --save_path exps/wavrgb/exp2

python trainSpeakerNet.py --model ResNetSE34Multi --modality wavrgbthr --noisy_train True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --max_epoch 600 --batch_size 100 --nPerSpeaker 9 --max_frames 200 --eval_frames 200 --weight_decay 0.1 --seed 1 --save_path exps/wavrgb/exp2

Evaluating pretrained models: clean test data

Unimodal models

python trainSpeakerNet.py --model ResNetSE34Multi --modality wav --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/wav/exp1

python trainSpeakerNet.py --model ResNetSE34Multi --modality rgb --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt   --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/rgb/exp1

python trainSpeakerNet.py --model ResNetSE34Multi --modality thr --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt   --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/thr/exp1

Multimodal models

python trainSpeakerNet.py --model ResNetSE34Multi --modality wavrgb  --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt   --log_input True  --trainfunc angleproto --eval_frames 200 --save_path exps/wavrgb/exp1

python trainSpeakerNet.py --model ResNetSE34Multi --modality wavrgbthr --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt   --log_input True  --trainfunc angleproto --eval_frames 200 --save_path exps/wavrgb/exp1

Evaluating pretrained models: noisy test data

Unimodal models

python revalidate.py --model ResNetSE34Multi --modality wav --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/wav/exp2

python revalidate.py --model ResNetSE34Multi --modality wav --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt    --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/wav/exp2

python revalidate.py --model ResNetSE34Multi --modality rgb --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/rgb/exp2

python revalidate.py --model ResNetSE34Multi --modality rgb --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt    --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/rgb/exp2

python revalidate.py --model ResNetSE34Multi --modality thr --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/thr/exp2

python revalidate.py --model ResNetSE34Multi --modality thr --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt    --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/thr/exp2

Multimodal models

python revalidate.py --model ResNetSE34Multi --modality wavrgb --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/wavrgb/exp2

python revalidate.py --model ResNetSE34Multi --modality wavrgb --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt    --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/wavrgb/exp2

python revalidate.py --model ResNetSE34Multi --modality wavrgbthr --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/wavrgbthr/exp2

python revalidate.py --model ResNetSE34Multi --modality wavrgbthr --eval True --valid_model True --test_path data/test --test_list data/metadata/test_list.txt    --noisy_eval True --p_noise 0.3 --snr 8 --log_input True --trainfunc angleproto --eval_frames 200 --save_path exps/wavrgb/exp2

This repository contains code and data for "On the Multimodal Person Verification Using Audio-Visual-Thermal Data"

Related tags

Overview

trimodal_person_verification

Dependencies

Dataset

Training examples : clean data

Training examples : noisy data

Evaluating pretrained models: clean test data

Evaluating pretrained models: noisy test data

Owner

ISSAI

Explaining Hyperparameter Optimization via PDPs

LLVM-based compiler for LightGBM gradient-boosted trees. Speeds up prediction by ≥10x.

Official repository of the AAAI'2022 paper "Contrast and Generation Make BART a Good Dialogue Emotion Recognizer"

Python scripts for performing 3D human pose estimation using the Mobile Human Pose model in ONNX.

GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. (CVPR 2021)

ELSED: Enhanced Line SEgment Drawing

PyImpetus is a Markov Blanket based feature subset selection algorithm that considers features both separately and together as a group in order to provide not just the best set of features but also the best combination of features

Clustering is a popular approach to detect patterns in unlabeled data

Code release for DS-NeRF (Depth-supervised Neural Radiance Fields)

An implementation of the proximal policy optimization algorithm

For the paper entitled ''A Case Study and Qualitative Analysis of Simple Cross-Lingual Opinion Mining''

Deep learning for spiking neural networks

Code for the Weighted, Accelerated and Restarted Primal-dual algorithm. This algorithm achieves stable linear convergence for reconstruction from undersampled noisy measurements under an approximate sharpness condition. See the paper for details.

A modular PyTorch library for optical flow estimation using neural networks

App for identification of various objects. Based on YOLO v4 tiny architecture

img2pose: Face Alignment and Detection via 6DoF, Face Pose Estimation

Pairwise Learning for Neural Link Prediction for OGB (PLNLP-OGB)

Pixel-level Crack Detection From Images Of Levee Systems : A Comparative Study

MRQy is a quality assurance and checking tool for quantitative assessment of magnetic resonance imaging (MRI) data.

This project helps to colorize grayscale images using multiple exemplars.