Attribute-Aware Convolutional Neural Networks for Facial Beauty Prediction
Authors: Luojun Lin, Lingyu Liang, Lianwen Jin, Weijie Chen
IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments to evaluate the performance and explore the properties of Aa Net and P-Aa Net mainly on the SCUT-FBP5500 benchmark dataset. Several detailed ablation studies on this benchmark are carried out to justify the effectiveness of our proposed networks. To further stress the superiority, we also compare our method with related methods on the SCUT-FBP5500 and SCUT-FBP datasets, and the results show that our method achieves stateof-the-art performance on these benchmarks. |
| Researcher Affiliation | Collaboration | Luojun Lin1 , Lingyu Liang1 , Lianwen Jin1 and Weijie Chen2 1School of Electronic and Information Engineering, South China University of Technology, China 2Hikvision Research Institute, China |
| Pseudocode | No | The paper describes the architecture and mathematical formulations but does not include pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statements about open-sourcing the code or links to a code repository. |
| Open Datasets | Yes | Most of our experiments are conducted on the SCUT-FBP5500 dataset [Liang et al., 2018], which contains 5500 facial images with diverse attributes (e.g., male/female, Asian/Caucasian) and diverse labels (e.g., facial landmarks, beauty scores). Additionally, we also conduct experiments on the SCUT-FBP dataset [Xie et al., 2015] which contains 500 facial images sampled from Asian female subject. |
| Dataset Splits | Yes | To ensure the effectiveness, five-folds cross validation is performed. The average results of the validations are reported below. |
| Hardware Specification | Yes | All the experiments are carried out on Caffe [Jia et al., 2014] with a NVIDIA Geforce GTX Titan X GPU. |
| Software Dependencies | Yes | All the experiments are carried out on Caffe [Jia et al., 2014] with a NVIDIA Geforce GTX Titan X GPU. |
| Experiment Setup | Yes | All the facial images (350 350) are resized to 256 256 firstly. Then a 224 224 crop and horizontal flipping are performed randomly, followed by perpixel rescale to [0,1] and mean value subtraction. In the following experiments, Aa Net takes facial images and their corresponding attributes as inputs, while the others use the facial images alone. For Alex Net and its related networks, they are trained by using mini-batch Stochastic Gradient Descent (SGD) with a batch size of 32, a momentum of 0.9, and a weight decay of 5e-4. We use a specific learning policy that the learning rate is increased from 0 to a peak value of 0.01 in a warm-up schedule of 2K iterations, and then decreased to 0, linearly, in 18K iterations. Note that, for P-Aa Net, the learning rate of pseudo attribute distiller is 0.1 times as that of the main network in order to ensure stable training. For Res Net-18 and its extension networks, we set the peak value of learning rate and weight decay as 0.1 and 1e-4. |