Multi-Scale Similarity Learning for Facial Age Estimation

This is the Pytorch implementation of the ICCAS 2023 paper, "Multi-scale Similarity Learning for Age Estimation based on Facial Images".

Paper

In this paper we propose a similarity learning model for age estimation, called Multi-scale Similarity Learning. We employ a multi-scale feature extraction to our model and prove its ability to extract more reliable features for a similarity learning. The proposed single framework architecture is trained end-to-end to simultaneously learn feature similarities and age differences. The experiments showed that the proposed approach provides superior age estimation performances.

Datasets

UTKFace
CACD

The folder datalists contains .csv files for separate datasets. The folder structure used for implementation:

./datalists/
    CACD/
        test.csv
        train.csv
        val.csv
        images/
            14_Aaron_Johnson_0001.jpg
            14_Aaron_Johnson_0002.jpg
            ...
    UTKFace/
        test.csv
        train.csv
        val.csv
        images/ 
            1_0_0_20161219140623097.jpg
            1_0_0_20161219140627985.jpg
            ...

You can put images under corresponding UTKFace and CACD folders or use your own paths to datasets. See in Evaluation and Training.

Dependencies

Python 3
Pytorch

Preprocessing

UTKFace. We use the aligned and cropped version of dataset.
CACD. We use the implementation of MTCNN by urbaneman to crop and align images.

Pretrained Models

The backbone we use is IR50, pre-trained on MS-Celeb-1M Dataset and modified by Talented-Q. You need to put the pretrained model under pretrained_models folder.

./pretrained_models/ir50.pth

Trained Models

The trained models for UTKFace and CACD datasets can be downloaded here.

In order to evaluate the trained models you can put them under trained_models folder or use your own paths to the models. See in Evaluation and Training.

Evaluation and Training

In opt.py file, you can configure the dataset type, the paths to images and trained model, the path to save the new checkpoints for model and other hyperparameters. Besides, you can configure the arguments explicitly in command line before running.

Testing

You can test the trained model on UTKFace dataset by running:

python Main.py --phase test --data_type utk --data_path [path/to/utk/images/ or ./datalists/UTKFace/images/] --model_path [path/to/trained/model or ./trained_models/utk.pt]

You can test the trained model on CACD dataset by running:

python Main.py --phase test --data_type cacd --data_path [path/to/cacd/images/ or ./datalists/CACD/images/] --model_path [path/to/trained/model or ./trained_models/cacd.pt]

Training and Validation

You can train and validate the model on UTKFace dataset by running:

python Main.py --phase train --data_type utk --data_path [path/to/utk/images/ or ./datalists/UTKFace/images/] --checkpoint_path path/to/save/checkpoints

You can train and validate the model on CACD dataset by running:

python Main.py --phase train --data_type cacd --data_path [path/to/cacd/images/ or ./datalists/CACD/images/] --checkpoint_path path/to/save/checkpoints --batch 64 --epochs 20 --seed 123

Citing

If you find this work useful, please cite the following paper:

@INPROCEEDINGS{10316890,
   author = {Amirullaeva, Sevara and Han, Ji-Hyeong},
   booktitle={2023 23rd International Conference on Control, Automation and Systems (ICCAS)}, 
   title = {Multi-Scale Similarity Learning for Age Estimation Based on Facial Images},
   year = {2023},
   pages={1874-1878},
   doi={10.23919/ICCAS59377.2023.10316890}
}

SeoulTech-HCIRLab/Multi-Scale-Age-Similarity-Learning