Searching for Alignment in Face Recognition
Authors: Xiaqing Xu, Qiang Meng, Yunxiao Qin, Jianzhu Guo, Chenxu Zhao, Feng Zhou, Zhen Lei3065-3073
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on our proposed benchmark validate the effectiveness of our method to improve face recognition performance. |
| Researcher Affiliation | Collaboration | 1AIBEE, Beijing, China, 2Northwestern Polytechnical University, Xian, China 3 CBSR & NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing, China 4 School of Artificial Intelligence, University of Chinese Academy of Sciences 5Academy of Sciences, Mininglamp Technology, Beijing, China |
| Pseudocode | Yes | Algorithm 1 Face Alignment Policy Search(FAPS). and Algorithm 2 The FAPS explore function. |
| Open Source Code | No | No explicit statement about providing open-source code for the described methodology or a link to a code repository was found. |
| Open Datasets | Yes | We separately employ CASIA (Yi et al. 2014) and MS-Celeb-1M (Guo et al. 2016) as middle-scale and large-scale training and searching datasets. |
| Dataset Splits | No | The validation set is designed considering the main challenges of face recognition like age, pose and illumination variations. As a result, we build a validation dataset named Cross Challenge in the Wild (CCW), the images are from three datasets in unconstrained environments: LFW(Huang et al. 2008), Age DB-30(Moschoglou et al. 2017) and CPLFW (Zheng and Deng 2018). (This describes the datasets used for validation, but does not provide specific details on how the training, validation, and test splits were performed for reproducibility, e.g., percentages, counts, or a specific splitting methodology for CCW itself.) |
| Hardware Specification | Yes | costs about 9102 GPU hours with 8 Tesla V100 GPUs. and FAPS s searching process takes 131 GPU hours with 8 Tesla V100 GPUs. |
| Software Dependencies | No | We implement FAPS with Py Torch (Paszke et al. 2019) and Ray Tune (Moritz et al. 2018). (Specific version numbers for PyTorch and Ray Tune are not provided.) |
| Experiment Setup | Yes | The embedding dimension is set to 512. To accelerate the searching process, Res Net18 is adopted as the searching network. Res Net50 is used to train on the training set. Arc Face (Deng et al. 2019) is served as the loss function during searching and training. ... During searching, the population size of models N is set to 8. The crop size mmax and mmin are set to 232 and 160, respectively. The vertical shift δmax and δmin are 24 and 32. We set the magnitude parameter of crop size sm = 8 and the magnitude parameter of vertical shift sδ = 4. and The learning rate is initialized by 0.1 and divided by 10 at epoch 20 and 28. |