Demodalizing Face Recognition with Synthetic Samples

Authors: Zhonghua Zhai, Pengju Yang, Xiaofeng Zhang, Maji Huang, Haijing Cheng, Xuejun Yan, Chunmao Wang, Shiliang Pu3278-3286

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the effectiveness of our approach on various benchmarks of large-scale face recognition and outperform the previous methods, especially in the low FAR range.
Researcher Affiliation Collaboration 1Hikvision Research Institute 2Zhejiang University
Pseudocode No The paper describes the proposed methods using textual descriptions and figures but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper mentions a GitHub link for a dataset built (CP-IJB-C) but does not explicitly state that the source code for the proposed DMFR methodology is open source or provide a link for it.
Open Datasets Yes We use a cleaned version of the MSCeleb-1M datasets (Guo et al. 2016) with 2,251,420 images of 58,982 subjects as our training set.
Dataset Splits No The paper mentions using a training set (MSCeleb-1M) and various test sets (LFW, YTF, IJB-A, etc.) but does not provide specific details on how the training data itself is partitioned into training and validation sets for model development.
Hardware Specification No The paper states: "We train the model with 8 synchronized graphic processing units (GPUs)" but does not specify the exact GPU models, CPU, or other hardware specifications used.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or other libraries).
Experiment Setup Yes We use an initial learning rate of 0.01 and reduce the learning rate by 0.1 at 50k, 70k and 80k with a weight decay of 0.0005 and a momentum of 0.9 using stochastic gradient descent (SGD). We train the model with 8 synchronized graphic processing units (GPUs) and a mini-batch, including 128 images per GPU. We empirically set λmeta = 0.5, λdisen = 1.0 and λflt = 1.0, respectively. The margin m is empirically set to 0.4.