Age Estimation

This is a PyTorch implementation of mean-variance loss [1] and softmax loss embedded into ResNet18 for age estimation.

Dependencies

• Python 3.7+
• PyTorch 1.0+
• Pillow 5.4+
• NumPy 1.16+
• TorchVision 0.2+
• OpenCV 4.1.1+

Tested on:

• Ubuntu 18.04, CUDA 10.1
• CPU: Intel Broadwell, GPU: NVIDIA Tesla K80

Usage

Setting

• I follow a widely used leave-one-person-out (LOPO) protocol in my experiments. Images of a person is used as test data, images of the others is used as training and validation data. Ratio of validation data to training and validation data is set as "VALIDATION_RATE= 0.1" in "main.py".

Training

1. Prepare a FG-NET dataset. DataLoaders of this program load images from a FG-NET dataset. A image filename of the FG-NET is prefixed by a subject index and contains a age label. Thus subject indices and labels are extracted from filenames.

URL: http://yanweifu.github.io/FG_NET_data/FGNET.zip

2. Start training

Mean-Variance Loss

python main.py --batch_size 64 --image_directory FGNET/images --leave_subject 1 --learning_rate 0.001 --epoch 100 --result_directory result/mean_variance -loss

--leave_subject: a subject index for test data (1~82). (integer)
--result_directory: a directory where the model will be saved.
-loss: mean-variance loss

Only Softmax Loss

python main.py --batch_size 64 --image_directory FGNET/images --leave_subject 1 --learning_rate 0.001 --epoch 100 --result_directory result/softmax

--leave_subject: a subject index for test data (1~82). (integer)
--result_directory: a directory where the model will be saved.

Comparison of mean-variance loss with softmax cross entropy

The following figure shows Mean Absolute Error (MAE) when the subject #1 is chosen as test data. We see that MAE on validation and test data of mean-variance loss and softmax cross entropy loss is better than that of softmax cross entropy loss only.

Inference

python main.py -pi FGNET/images/001A43a.JPG -pm result/model_best_loss

The following figures show the results of the subject #1 who I choosed as test data. The left image is the picture when the subject was 2 years old. The right image is the picture when the subject was 43 years old.

Differences between this implementation and original a paper.

• In the original paper, AlexNet and VGG-16 are used as feature extractors. On the other hand, ResNet18 is used in this program.
• In the original paper, SGD optimizer is used. On the other hand, Adam optimizer is used in this program.
• In the original paper, all the face images are aligned based on five facial landmarks detected using an open-source SeetaFaceEngine [2]. In this program, no alignment is done.
• In this program, three augmentations are adopted to images randomly, i.e., horizontal flip, rotate and shear.

References

[1] H. Pan, et al. "Mean-Variance Loss for Deep Age Estimation from a Face." Proceedings of CVPR, 2018.

[2] https://github.com/seetaface/SeetaFaceEngine