Hierarchical Multi-task Network For Race, Gender and Facial Attractiveness Recognition

Introduction

This repository holds the PyTorch implementation of our paper Hierarchical Multi-Task Network For Race, Gender and Facial Attractiveness Recognition.

Updated

1. By leveraging FiveCrops inference, we are able to achieve better performance.
2. We also report 5 cross validation results, since we find newly proposed models often use this metric instead of 6/4 splitting strategy.

How to use

• Install 3rd party libraries
  sudo pip3 install -r requirements.txt
• Modify cfg.py to fit your path

Hyper-param Selection

Loss	MAE	RMSE	PC	Acc_R	Acc_G	Epoch	WD
MSE	0.2556	0.3372	0.8693	99.68%	98.53%	170	5e-2
L1	0.2500	0.3299	0.8753	99.26%	98.16%	150	5e-2
Smooth L1	0.2531	0.3313	0.8738	99.54%	98.58%	170	5e-2
Smooth Huber	0.2501	0.3263	0.8783	99.26%	98.16%	170	5e-2
Smooth Huber + FiveCrops (New)	0.2439	0.3226	0.8801	99.45%	98.58%	149	5e-2

Performance Comparison

6/4 Splitting

Methods	MAE	RMSE	PC
ResNeXt-50	0.2518	0.3325	0.8777
ResNet-18	0.2818	0.3703	0.8513
AlexNet	0.2938	0.3819	0.8298
CRNet	0.2816	0.3669	0.8450
HMTNet (Ours)	0.2500	0.3299	0.8753
HMTNet (Ours)	0.2501	0.3263	0.8783
HMTNet (Ours)	0.2439	0.3226	0.8801

5 Fold Cross Validation

5-fold CV Results

Round	Acc_r	Acc_g	MAE	RMSE	PC
1	99.54%	98.53%	0.2357	0.3091	0.8915
2	99.72%	98.62%	0.2365	0.3150	0.8884
3	99.54%	99.17%	0.2442	0.3235	0.8863
4	99.63%	98.16%	0.2335	0.3053	0.9006
5	99.36%	99.26%	0.2403	0.3178	0.8892
Avg	99.56%	98.75%	0.2380	0.3141	0.8912

Comparison with others

Methods	MAE	RMSE	PC
ResNeXt-50	0.2291	0.3017	0.8997
ResNet-18	0.2419	0.3166	0.8900
AlexNet	0.2651	0.3481	0.8634
HMTNet	0.2380	0.3141	0.8912

Samples

Deep Feature Visualization

Resources

• Pretrained HMTNet

Citation

If you find this repository helps your research, please cite our paper:

@inproceedings{xu2019hierarchical,
  title={Hierarchical Multi-Task Network For Race, Gender and Facial Attractiveness Recognition},
  author={Xu, Lu and Fan, Heng and Xiang, Jinhai},
  booktitle={2019 IEEE International Conference on Image Processing (ICIP)},
  pages={3861--3865},
  year={2019},
  organization={IEEE}
}