Impact of Image Resolution on Age Estimation with DeepFace and InsightFace

DeepFace vs. InsightFace

This repository contains the materials related to a study on how input image resolution affects the accuracy and processing performance of age estimation using DeepFace and InsightFace. The project explores seven different resolutions applied to 1000 images from the IMDB-Clean dataset, resulting in 7000 processed samples.

---

Overview

Automatic age estimation is used in applications such as age verification, access control and automated checkout systems. Input images in these scenarios vary greatly in quality and resolution. This study evaluates how different input resolutions influence the accuracy of two widely used frameworks and whether external resolution still matters despite internal resizing.

This study compares:

• DeepFace (VGG-Face)
• InsightFace (ArcFace + genderage head)
• Seven resolutions from 64×64 up to 1080×1080
• 7000 total processed samples

---

Key Findings

• Both DeepFace and InsightFace achieve their best performance at 224×224 pixels.
• At 224×224, the results were:
  • DeepFace: MAE 10.83 years
  • InsightFace: MAE 7.46 years
• Low resolutions (e.g., 64×64) lead to substantially higher errors.
• Very high resolutions (e.g., 1080×1080) also reduce accuracy.
• InsightFace is significantly faster than DeepFace at all resolutions, typically between 0.015–0.02 seconds per image.

---

Repository Contents

File	Description
paper.pdf	English version of the research paper
poster.pdf	Visual research poster
paper-nl.pdf	Dutch version of the research paper
Colab Notebook	Code used for preprocessing, testing and analysis
Wil je dat ik deze ook in de Nederlandse README of Engelse README integreer?

---

Google Colab Notebook

The complete experimental pipeline, including preprocessing, inference and metric calculations, is available in this Colab notebook:

https://colab.research.google.com/drive/1roSZ2Y4Ne9Yy71DM2mRmY0VetfuydbEi

---

Method Summary

• Dataset: IMDB-Clean (1000 selected images)
• Resolutions tested: 64×64, 112×112, 224×224, 256×256, 512×512, 720×720, 1080×1080
• Frameworks: DeepFace (VGG-Face), InsightFace (ArcFace + genderage head)
• Metrics: MAE, Standard Deviation, Median Absolute Error
• Hardware: Google Colab T4 GPU

---

Conclusion

Input image resolution has a clear and consistent effect on age estimation accuracy. Both frameworks perform best at 224×224 pixels, and InsightFace achieves higher accuracy and faster inference overall. For real-time applications, InsightFace at 224×224 is recommended.

---

📜 Citation

If you use this repository, data, or results in your research, please cite the following paper:

BibTeX:

@article{jamo2025resolutionageestimation,
  title   = {Impact of Image Resolution on Age Estimation with DeepFace and InsightFace},
  author  = {Jamo, Shiyar},
  journal = {arXiv preprint arXiv:2511.14689},
  year    = {2025}
}

📄 Paper:
https://arxiv.org/abs/2511.14689

🔗 View PDF:
https://arxiv.org/pdf/2511.14689

🔗 View HTML:
https://arxiv.org/html/2511.14689v1

Author

Shiyar Jamo
Master Applied Artificial Intelligence
Amsterdam University of Applied Sciences

Figures

Figure 1 — Google Colab Environment Setup

This figure shows the runtime configuration used for all experiments.
A T4 GPU was selected in Google Colab to ensure consistent hardware performance across all tests.
Both DeepFace and InsightFace were executed using Python 3 with the latest stable runtime version.

---

Figure 2 — Average Inference Time per Resolution

This figure illustrates the mean inference time of DeepFace and InsightFace across seven input resolutions, from 64×64 to 1080×1080 pixels.

• InsightFace remains consistently fast (≈0.015–0.02 s).
• DeepFace shows a strong increase in processing time as resolution grows, reaching 0.7 s at 1080×1080.
  These results highlight the greater computational efficiency of InsightFace.

---

Figure 3 — Error Metrics per Resolution (MAE, SD, MedAE)

This composite figure presents three key error metrics per resolution:

• MAE (Mean Absolute Error)
• Standard Deviation
• Median Absolute Error

InsightFace consistently outperforms DeepFace across all metrics, with the lowest errors occurring at 224×224 pixels.
Both very low and very high resolutions lead to degraded performance.

Repository

https://github.com/codershiyar/image-resolution-effect-on-ai