Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FRoundation: Are Foundation Models Ready for Face Recognition?

Tahar Chettaoui, Naser Damer, Fadi Boutros

TL;DR

This paper investigates whether vision foundation models can be repurposed for face recognition (FR) and how to adapt them when data is scarce or synthetic data is used. It introduces FRoundation, a LoRA-based fine-tuning framework applied to DINOv2 and CLIP ViTs, paired with CosFace loss for FR, and evaluates on standard FR benchmarks with varying data availability. Key findings show that pre-trained foundation models underperform FR-specific models, but fine-tuning with LoRA yields substantial gains, especially with limited data (e.g., average accuracy rising from $64.70\%$ to $87.10\%$ with 1K identities), while large-scale training from scratch can be competitive but far more costly; synthetic data further enhances performance beyond both pre-trained and baseline models. The results underscore the practical potential of foundation models for FR in data-constrained settings, while also noting bias considerations and the value of synthetic data to broaden FR readiness.

Abstract

Foundation models are predominantly trained in an unsupervised or self-supervised manner on highly diverse and large-scale datasets, making them broadly applicable to various downstream tasks. In this work, we investigate for the first time whether such models are suitable for the specific domain of face recognition (FR). We further propose and demonstrate the adaptation of these models for FR across different levels of data availability, including synthetic data. Extensive experiments are conducted on multiple foundation models and datasets of varying scales for training and fine-tuning, with evaluation on a wide range of benchmarks. Our results indicate that, despite their versatility, pre-trained foundation models tend to underperform in FR in comparison with similar architectures trained specifically for this task. However, fine-tuning foundation models yields promising results, often surpassing models trained from scratch, particularly when training data is limited. For example, after fine-tuning only on 1K identities, DINOv2 ViT-S achieved average verification accuracy on LFW, CALFW, CPLFW, CFP-FP, and AgeDB30 benchmarks of 87.10%, compared to 64.70% achieved by the same model and without fine-tuning. While training the same model architecture, ViT-S, from scratch on 1k identities reached 69.96%. With access to larger-scale FR training datasets, these performances reach 96.03% and 95.59% for the DINOv2 and CLIP ViT-L models, respectively. In comparison to the ViT-based architectures trained from scratch for FR, fine-tuned same architectures of foundation models achieve similar performance while requiring lower training computational costs and not relying on the assumption of extensive data availability. We further demonstrated the use of synthetic face data, showing improved performances over both pre-trained foundation and ViT models.

FRoundation: Are Foundation Models Ready for Face Recognition?

TL;DR

This paper investigates whether vision foundation models can be repurposed for face recognition (FR) and how to adapt them when data is scarce or synthetic data is used. It introduces FRoundation, a LoRA-based fine-tuning framework applied to DINOv2 and CLIP ViTs, paired with CosFace loss for FR, and evaluates on standard FR benchmarks with varying data availability. Key findings show that pre-trained foundation models underperform FR-specific models, but fine-tuning with LoRA yields substantial gains, especially with limited data (e.g., average accuracy rising from to with 1K identities), while large-scale training from scratch can be competitive but far more costly; synthetic data further enhances performance beyond both pre-trained and baseline models. The results underscore the practical potential of foundation models for FR in data-constrained settings, while also noting bias considerations and the value of synthetic data to broaden FR readiness.

Abstract

Foundation models are predominantly trained in an unsupervised or self-supervised manner on highly diverse and large-scale datasets, making them broadly applicable to various downstream tasks. In this work, we investigate for the first time whether such models are suitable for the specific domain of face recognition (FR). We further propose and demonstrate the adaptation of these models for FR across different levels of data availability, including synthetic data. Extensive experiments are conducted on multiple foundation models and datasets of varying scales for training and fine-tuning, with evaluation on a wide range of benchmarks. Our results indicate that, despite their versatility, pre-trained foundation models tend to underperform in FR in comparison with similar architectures trained specifically for this task. However, fine-tuning foundation models yields promising results, often surpassing models trained from scratch, particularly when training data is limited. For example, after fine-tuning only on 1K identities, DINOv2 ViT-S achieved average verification accuracy on LFW, CALFW, CPLFW, CFP-FP, and AgeDB30 benchmarks of 87.10%, compared to 64.70% achieved by the same model and without fine-tuning. While training the same model architecture, ViT-S, from scratch on 1k identities reached 69.96%. With access to larger-scale FR training datasets, these performances reach 96.03% and 95.59% for the DINOv2 and CLIP ViT-L models, respectively. In comparison to the ViT-based architectures trained from scratch for FR, fine-tuned same architectures of foundation models achieve similar performance while requiring lower training computational costs and not relying on the assumption of extensive data availability. We further demonstrated the use of synthetic face data, showing improved performances over both pre-trained foundation and ViT models.

Paper Structure

This paper contains 22 sections, 4 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Foundation models:
  4. Methodology
  5. Baseline Foundation Models
  6. DINOv2
  7. CLIP
  8. Fine-tuning with LoRA
  9. FRoundation
  10. Experimental setups
  11. Evaluation Datasets
  12. Training and Fine-tuning Datasets
  13. Model architectures
  14. Training Settings
  15. Computational Cost
  16. ...and 7 more sections

Figures (3)

  • Figure 1: ViT Training Pipeline with LoRA Adapters. The facial image is divided into patches, which are mapped to patch embeddings via a linear projection. A class token is added, and positional embeddings are incorporated to maintain spatial information. This sequence of embedding vectors is then input into the encoder. The transformer layers remain fixed during training, while trainable LoRA layers are introduced to fine-tune the model. Each LoRA layer operates independently within the transformer layers and possesses its own set of weights.
  • Figure 2: LoRA Integration in Multi-Headed Self-Attention Block. We implement LoRA to $q$ and $v$ projection layers in each transformer block.
  • Figure 3: Average verification accuracies on five benchmarks, LFW, CFP-FP, AgeDB30, CALFW, and CPLFW on the y-axis vs. training/fine-tuning dataset size, in terms of the number of identities, on the x-axis. The results correspond to the ones reported in Table \ref{['table:portion_data']} and plotted for ViT (baseline) as well as fine-tuned DINOv2 and CLIP. Increasing the training/fine-tuning dataset width (number of identities) improved the model recognition performances.