SphereFace2: Binary Classification is All You Need for Deep Face Recognition
Authors: Yandong Wen, Weiyang Liu, Adrian Weller, Bhiksha Raj, Rita Singh
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on popular benchmarks demonstrate that Sphere Face2 can consistently outperform state-of-the-art deep face recognition methods. |
| Researcher Affiliation | Academia | 1Carnegie Mellon University 2University of Cambridge 3MPI for Intelligent Systems 4Alan Turing Institute |
| Pseudocode | No | The paper describes mathematical formulations and algorithms in text and equations, but it does not provide any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at Open Sphere. |
| Open Datasets | Yes | We adopt VGGFace2 [2] as the same training set for all the methods. VGGFace2 contains 3.1M images from 8.6K identities, as shown in Table 9. We perform an ablation study on four validation sets: LFW, Age DB-30, CA-LFW, and CP-LFW. The statistics of these datasets are summarized in Table 9. |
| Dataset Splits | Yes | We perform an ablation study on four validation sets: LFW, Age DB-30, CA-LFW, and CP-LFW. The statistics of these datasets are summarized in Table 9. Following the provided evaluation protocols, we report 1:1 verification accuracy of 6,000 pairs (3,000 positive and 3,000 negative pairs) for each dataset. In addition, we combine these datasets and compute the overall verification accuracy, which serves as a more accurate metric to evaluate the models. |
| Hardware Specification | No | The paper mentions using "GPUs" and discusses "multi-GPU model parallelization" (e.g., "Fig. 10 (right) shows how the number of processed images per second changes with different numbers of GPUs."), but it does not specify any particular GPU models, CPU models, memory, or other detailed hardware specifications used for the experiments. |
| Software Dependencies | No | The paper mentions using MTCNN [48] for face landmark detection but does not provide specific version numbers for any software dependencies used in the experiments (e.g., deep learning frameworks like PyTorch/TensorFlow, or other libraries with their versions). |
| Experiment Setup | Yes | We use SGD with momentum 0.9 by default. We adopt VGGFace2 [2] as the same training set for all the methods. The training faces are horizontally flipped for data augmentation. We use 20-layer CNNs in ablations and 64-layer CNNs for the comparison to existing state-of-the-art methods. Since r and m have been extensively explored in [4, 37 39], we follow previous practice to fix r and m to 30 and 0.4 respectively. Our experiments mainly focus on analyzing the effect of λ and t. We use three challenging face recognition benchmarks, IJB-B, IJB-C, and Mega Face to evaluate Sphere Face2 (with λ=0.7, r=40, m=0.4, t=3). |