HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere
Authors: Hatef Otroshi Shahreza, Sébastien Marcel
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results show that models trained with Hyper Face achieve state-of-the-art performance in training face recognition using synthetic datasets. Project page: https://www.idiap.ch/paper/hyperface |
| Researcher Affiliation | Academia | Hatef Otroshi Shahreza1,2 and S ebastien Marcel1,3 1Idiap Research Institute, Martigny, Switzerland 2 Ecole Polytechnique F ed erale de Lausanne (EPFL), Lausanne, Switzerland 3Universit e de Lausanne (UNIL), Lausanne, Switzerland EMAIL |
| Pseudocode | Yes | Algorithm 1 Hyper Face Optimization for Finding Reference Embeddings Algorithm 2 Hyper Face Stochastic Optimization for Finding Reference Embeddings Algorithm 3 Hyper Face Dataset Generation |
| Open Source Code | Yes | The source code of our experiments and generated datasets are publicly available3. 3Project page: https://www.idiap.ch/paper/hyperface |
| Open Datasets | Yes | We evaluate the recognition performance of models trained using synthetic datasets, and show that our optimization and packing approach can lead to new synthetic datasets that can be used to train face recognition models. We also compare trained models with our generated dataset to models trained with previous synthetic datasets, where our generated datasets achieve competitive performance with state-of-the-art synthetic datasets in the literature. Figure 1 illustrates sample face images from our synthetic dataset. ... benchmarking datasets of real images, including Labeled Faces in the Wild (LFW) (Huang et al., 2008), Cross-age LFW (CA-LFW) (Zheng et al., 2017), Cross Pose LFW (CP-LFW) (Zheng & Deng, 2018), Celebrities in Frontal-Profile in the Wild (CFP-FP) (Sengupta et al., 2016), and Age DB-30 (Moschoglou et al., 2017) datasets. |
| Dataset Splits | Yes | For consistency with prior works, we report recognition accuracy calculated using 10-fold cross-validation for each of benchmarking datasets. |
| Hardware Specification | Yes | For example, we can solve the optimization for n X = 512 and nid = 10, 000 on a system equipped with a single NVIDIA 3090 GPU in 6 hours. |
| Software Dependencies | No | The paper mentions several software tools and models, such as Arc Face (Deng et al., 2019), Adam optimizer (Kingma & Ba, 2015), i Res Net50 backbone, Ada Face loss function (Kim et al., 2022), Stochastic Gradient Descent (SGD), Style GAN (Karras et al., 2020), Latent Diffusion Model (LDM) (Rombach et al., 2022), and MTCNN (Zhang et al., 2016). However, specific version numbers for these tools or for general programming languages/libraries (like Python, PyTorch) are not provided. |
| Experiment Setup | Yes | Dataset Generation: For solving the Hyper Face optimization in Algorithm 1, we use an initial learning rate of λ = 0.01 and reduce the learning rate by power 0.75 every 5, 000 iterations for a total number of iterations nitr = 100, 000. We also consider cosine distance, which is commonly used in face recognition systems for the comparison of face embeddings, as our distance function d( , ). For the hyperparameters α and β, we consider default values of 0.5 and 0.01, respectively, in our experiments. We also consider the size of gallery to be the same as the number of identities, and explore other cases where ngallery > nid in our ablation study. We generate 64 images, by default, per each identity in our generated datasets and explore other numbers of images in our ablation study. Evaluation: To evaluate the generated synthetic datasets, we use each generated datasets as a training dataset for training a face recognition model. To this end, we use the i Res Net50 backbone and train the model with Ada Face loss function (Kim et al., 2022) using the Stochastic Gradient Descent (SGD) optimizer with the initial learning rate 0.1 and a weight decay of 5 10 4 for 30 epochs with the batch size of 256. |