Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ef-QuantFace: Streamlined Face Recognition with Small Data and Low-Bit Precision

William Gazali, Jocelyn Michelle Kho, Joshua Santoso, Williem

TL;DR

This paper demonstrates that effective quantization is achievable with a smaller dataset, presenting a new paradigm, and incorporates an evaluation-based metric loss and achieves an outstanding 96.15% accuracy on the IJB-C dataset, establishing a new state-of-the-art compressed model training for face recognition.

Abstract

In recent years, model quantization for face recognition has gained prominence. Traditionally, compressing models involved vast datasets like the 5.8 million-image MS1M dataset as well as extensive training times, raising the question of whether such data enormity is essential. This paper addresses this by introducing an efficiency-driven approach, fine-tuning the model with just up to 14,000 images, 440 times smaller than MS1M. We demonstrate that effective quantization is achievable with a smaller dataset, presenting a new paradigm. Moreover, we incorporate an evaluation-based metric loss and achieve an outstanding 96.15% accuracy on the IJB-C dataset, establishing a new state-of-the-art compressed model training for face recognition. The subsequent analysis delves into potential applications, emphasizing the transformative power of this approach. This paper advances model quantization by highlighting the efficiency and optimal results with small data and training time.

Ef-QuantFace: Streamlined Face Recognition with Small Data and Low-Bit Precision

TL;DR

This paper demonstrates that effective quantization is achievable with a smaller dataset, presenting a new paradigm, and incorporates an evaluation-based metric loss and achieves an outstanding 96.15% accuracy on the IJB-C dataset, establishing a new state-of-the-art compressed model training for face recognition.

Abstract

In recent years, model quantization for face recognition has gained prominence. Traditionally, compressing models involved vast datasets like the 5.8 million-image MS1M dataset as well as extensive training times, raising the question of whether such data enormity is essential. This paper addresses this by introducing an efficiency-driven approach, fine-tuning the model with just up to 14,000 images, 440 times smaller than MS1M. We demonstrate that effective quantization is achievable with a smaller dataset, presenting a new paradigm. Moreover, we incorporate an evaluation-based metric loss and achieve an outstanding 96.15% accuracy on the IJB-C dataset, establishing a new state-of-the-art compressed model training for face recognition. The subsequent analysis delves into potential applications, emphasizing the transformative power of this approach. This paper advances model quantization by highlighting the efficiency and optimal results with small data and training time.
Paper Structure (27 sections, 4 equations, 2 figures, 5 tables)

This paper contains 27 sections, 4 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Face Recognition
  4. Model Compression
  5. Model Quantization.
  6. Model Quantization in FR
  7. Proposed Method
  8. Efficient Quantization with EKD loss
  9. Experiments
  10. Dataset
  11. Prior Work Training Set.
  12. Our Training Set.
  13. Test Set.
  14. Experimental setting
  15. Model Architecture.
  16. ...and 12 more sections

Figures (2)

  • Figure 1: The illustration of our proposed solution compared to QuantFace BoutrosICPR2022. Our method efficiently reduced the training time by 440$\times$ while still achieving state-of-the-art performance. Bold text represents the best score.
  • Figure 2: An overview of our proposed pipeline. The compression of the teacher network is executed through either quantization or pruning techniques. After the model compression step, the fine-tuning process is initiated under a knowledge distillation paradigm, with information distilled from the final feature embedding layer.