lucasxlu / transfbp

Official Code for Paper <Transferring Rich Deep Features for Facial Beauty Prediction> (arXiv1803.07253)

Transferring Rich Deep Features for Facial Beauty Prediction

Introduction

This repo provides the source code for our paper Transferring Rich Deep Features for Facial Beauty Prediction. This code has been tested on Ubuntu16 .04 with TensorFlow0.12.0, a newer version may bring you some trouble since TensorFlow's APIs always change after releasing a new version.

Proposed Method

Experiments

Our proposed two-stage method achieves state-of-the-art performance on SCUT-FBP and Female Facial Beauty Dataset (ECCV2010) v1.0 dataset. TransFBP also achieves very competitive performance on SCUT-FBP5500 dataset.

• Evaluation with the SCUT-FBP Dataset

Methods	PC
Combined Features+Gaussian Reg	0.6482
CNN-based	0.8187
Liu et al.	0.6938
KFME	0.7988
RegionScatNet	0.83
PI-CNN	0.87
TransFBP (Ours)	0.8742

• Evaluation with the HotOrNot Dataset

Methods	PC
Eigenface	0.180
Multiscale Model	0.458
Auto Encoder	0.437
TransFBP (Ours)	0.468

• Evaluation with the SCUT-FBP5500 Dataset

Methods	PC
Geometric features + Gaussian Regression	0.6738
Geometric features + SVR	0.6668
64UniSample + SVR	0.8065
AlexNet	0.8298
ResNet18	0.8513
ResNeXt50	0.8777
TransFBP (Ours)	0.8519

Examples

Resources

• Pretrained Model on SCUT-FBP5500
• Pretrained VGG-Face Weights

Citation

If you find the code or the experimental results useful in your research, please consider citing our paper as:

@article{xu2018transferring,
  title={Transferring Rich Deep Features for Facial Beauty Prediction},
  author={Xu, Lu and Xiang, Jinhai and Yuan, Xiaohui},
  journal={arXiv preprint arXiv:1803.07253},
  year={2018}
}

transfbp's Issues

Sizing issue in inference using pre-trained models

Hey @lucasxlu
I tried following your instructions, however I am facing some issues in inferencing the pre-trained models.
I kept the infer.py almost same, but just changed paths for vgg-face.mat and used joblib to load the BayesRegression model. However I face the following error -

ValueError: matmul: Input operand 1 has a mismatch in its core dimension 0, with gufunc signature (n?,k),(k,m?)->(n?,m?) (size 8192 is different from 100352)

I think you have trained the model using some other layer(and not conv5_1) or maybe 2 FC layers concatenated with each other(4096+4096)?

Pretrained models for inference

Hey @lucasxlu ,

Thanks a lot for providing amazing results on the benchmark datasets. I was wondering on how to inference the TransFBP results on custom images. Is there a link to the pre-trained model available? As well as any instructions for inference available?