Physically-Based Face Rendering for NIR-VIS Face Recognition
Authors: Yunqi Miao, Alexandros Lattas, Jiankang Deng, Jungong Han, Stefanos Zafeiriou
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on four challenging NIR-VIS face recognition benchmarks demonstrate that the proposed method can achieve comparable performance with the state-of-the-art (SOTA) methods without requiring any existing NIR-VIS face recognition datasets. |
| Researcher Affiliation | Collaboration | Warwick Manufacturing Group University of Warwick Yunqi.Miao.1@warwick.ac.uk Alexandros Lattas Imperial College London and Huawei a.lattas@imperial.ac.uk Jiankang Deng Huawei and Insight Face j.deng16@imperial.ac.uk Jungong Han Department of Computer Science Aberystwyth University jungonghan77@gmail.com Stefanos Zafeiriou Department of Computing Imperial College London s.zafeiriou@imperial.ac.uk |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and pretrained models are released under the insightface2 Git Hub. 2https://github.com/deepinsight/insightface/tree/master/recognition |
| Open Datasets | Yes | Four NIR-VIS face recognition datasets are used to evaluate the proposed method. Specifically, the CASIA NIR-VIS 2.0 [32] (725 identities) and the LAMP-HQ [52] (573 identities)... To acquire the 3D facial assets, we use Avatar Me++[30], trained with Real Face DB [29]... The dataset used is Celeb A [35]... we use a subset from the large-scale VIS dataset [56], i.e., Web Face4M [57]... |
| Dataset Splits | Yes | Following [32, 52], the ten-fold experiments are conducted on both datasets. For each fold, approximately 50% identities are randomly selected as the training set and the rest are adopted as the testing set. Note that, there is no overlap between the training and testing sets. In terms of two low-shot NIR-VIS face recognition datasets: the Oulu-CASIA NIR-VIS [26] and the BUAA-Vis Nir [25], the identities within the datasets are split as 20/20 and 50/100, respectively, for the setting of the training/testing set. |
| Hardware Specification | No | The paper mentions using 'Marmoset Toolbag' for rendering but does not specify any hardware details such as GPU/CPU models, memory, or specific computing environments used for experiments. |
| Software Dependencies | No | The paper mentions using 'Avatar Me++', 'GANFIT', 'Marmoset Toolbag', and 'Light CNN-29' as software/models and 'SGD' as an optimizer, but does not provide specific version numbers for any of these components. |
| Experiment Setup | Yes | During training, we first train the network with identity loss Lid for 20 epochs on the Web Face4M and the synthesized dataset. Then, we fine-tune the network on the synthesized images with both identity loss Lid and ID-MMD loss Lidmmd for 5 epochs. The batch size is set as 512. During fine-tuning, 32 identities are randomly selected to form a mini-batch, and for each identity, 8 VIS and 8 NIR images are randomly selected. Stochastic gradient descent (SGD) is adopted as the optimizer, where the momentum is set to 0.9 and the weight decay is set to 1e-4. The learning rate is set to 1e-2 initially and decays by a factor of 0.5 per 10 epochs. |