Relative Age Position Learning for Facial Age Estimation

This is the Pytorch implementation of the paper "Relative Age Position Learning for Face-based Age Estimation".

Paper

This paper introduces the concept of relative age position learning to supplement the learning process for age estimation models. Drawing inspiration from feature re-calibration modules that prioritize features based on their importance, a novel age-based re-weighting module is developed to enhance the feature representation in the proposed age estimation method. The proposed re-weighting module obtains the features of selected references for each age and further exploits them to re-weight the features of input images based on the age importance. The resulting re-calibrated features are then assessed for relative age position prediction. Through extensive experiments, we demonstrate that the proposed approach outperforms other state-of-the-art age estimation methods on three challenging benchmark datasets for facial age estimation such as AgeDB, AFAD, and CACD.

Datasets

• AgeDB
• AFAD
• CACD

The folder datalists contains .csv files for each dataset. The folder structure used for implementation:

./datalists/
    afad.csv
    agedb.csv
    cacd.csv

The fold columns in .csv files identify train, valid, test splits of datasets. You can use your own paths to dataset images. See in Evaluation and Training.

Dependencies

• Python 3
• Pytorch

Preprocessing

• AgeDB. The original images of dataset are used.
• AFAD. The original images of dataset are used.
• 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. We pre-trained the IR50 again on IMDB-Clean dataset. We provide the pretrained checkpoints for IR50 here. You need to put the pretrained model under pretrained_model folder.

./pretrained_model/ir50_imdb.pt

Trained Models

The trained checkpoints for AgeDB, AFAD and CACD datasets can be downloaded here.

In order to evaluate the trained models you can put them under trained_checkpoints 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 phase type (train or test), dataset name, the paths to images, the path to trained model or to save 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 AgeDB dataset by running:

python main.py --phase test --data_name agedb --data_path path/to/agedb/images/

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

python main.py --phase test --data_name afad --data_path path/to/afad/images/ --model_path ./trained_checkpoints/afad.pt

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

python main.py --phase test --data_name cacd --data_path path/to/cacd/images/ --model_path [./trained_checkpoints/cacd.pt or own/path/to/cacd/checkpoints]

Training and Validation

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

python main.py --phase train --data_name agedb --data_path path/to/agedb/images/ --model_path path/to/save/checkpoints

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

python main.py --phase train --data_name afad --data_path path/to/afad/images/ --model_path path/to/save/checkpoints --seed 123

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

python main.py --phase train --data_name cacd --data_path path/to/cacd/images/ --model_path path/to/save/checkpoints --lr 1e-6

Citing

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

@ARTICLE{10646358,
  author={Amirullaeva, Sevara and Han, Ji-Hyeong},
  journal={IEEE Access}, 
  title={Relative Age Position Learning for Face-Based Age Estimation}, 
  year={2024},
  volume={12},
  pages={118832-118841},
  doi={10.1109/ACCESS.2024.3445178}}