USING SYNTHETIC DATA FOR FACE ATTRIBUTES RECOGNITION

USING SYNTHETIC DATA FOR FACE ATTRIBUTES RECOGNITION
Raman Charapennikau
Belarussian State University
Abstract: The paper explores the usage of diffusion models to create synthetic datasets for facial attribute recognition, focusing on age, gender and ethnicity prediction. We compare models trained on real-world data, synthetic data, and a combination of both. We demonstrate that pretraining on synthetic data followed by fine-tuning on real samples outperforms models trained solely on real-world data. Our results highlight the potential of synthetic data to enhance neural network performance in regression and classification tasks.

Dataset Creation

First, Download Humans model from CivitAI and put it to pretrained_models folder.

Then run following command:

python create_dataset.py --dest-folder <DATASET_SAVE_FOLDER> --total-images <NUMBER_OF_IMAGES_TO_GENERATE> --num_images_per_folder <NUMBER_OF_IMAGES_TO_GENERATE> 

Model Training

UTK/Ours

For example to train on synthetic data only:

python train.py -c ./configs/synth-only.yaml

FairFace

For example to train on FairFace data only:

python ff_train.py -c ./configs/ff-only.yaml

Model Evaluation

Run following command

python eval.py --checkpoint <PATH_TO_CHECKPOINT>