More Information Supervised Probabilistic Deep Face Embedding Learning
Authors: Ying Huang, Shangfeng Qiu, Wenwei Zhang, Xianghui Luo, Jinzhuo Wang
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on several benchmarks demonstrate that LATSE help face embedding to gain more generalization capability and it boost the single model performance with open training dataset to more than 99% on Mega Face test. 4. Experiments |
| Researcher Affiliation | Collaboration | 1Guangzhou Huya Technology Co., Ltd, Guangzhou, Guangdong, China 2Department of Engineering Science, University of Oxford, Oxford, Oxfordshire, United Kingdom. |
| Pseudocode | Yes | Algorithm 1 Linear-Auto-TS-Encoder Learning algorithm |
| Open Source Code | No | The paper does not provide an explicit statement or link indicating that the source code for the described methodology is open-source or publicly available. |
| Open Datasets | Yes | We employed CASIA (Yi et al., 2014) as small training set and MS1MV2 (Guo et al., 2016) or MS1MRetina V (Deng et al., 2019a) from Arc Face (Deng et al., 2019a) as large training dataset. |
| Dataset Splits | No | The paper mentions training and testing datasets, and provides some "veriļ¬cation results" (e.g., Table 1), but it does not specify the splits (percentages or counts) for validation datasets or how they were created from the main datasets. |
| Hardware Specification | Yes | Models were trained in eight NVIDIA Tesla V100 GPUs(16GB) with total batch size 768. |
| Software Dependencies | No | The proposed method were implemented with MXNet (Chen et al., 2015). The paper mentions MXNet but does not provide specific version numbers for MXNet or any other software libraries or dependencies used. |
| Experiment Setup | Yes | The learning rate started from 0.1 and was divide by 10 at 10K, 16K, 20K, 22k iterations. We set weight decay to 5e 4 and momentum to 0.9. At test time, we only computed the 512 dimension feature for each normalized face from the student network and compared feature cosine angle value as the similarity score between different face images. |