Black and white crayon drawing of a research lab
Research

ICI Innolabs at NeurIPS 2025: Advancing Fairness in AI Face Detection

Bogdan Dura and Sebastian Balmus

We’re proud to announce that the ICI Innolabs team achieved 9th place in the NeurIPS 2025 Competition on Fairness in AI Face Detection — a global challenge focused on building accurate and equitable AI systems for face detection.

This competition, part of the NeurIPS (Conference on Neural Information Processing Systems) — one of the world’s leading venues in artificial intelligence and machine learning — brought together 162 registered participants and over 2,100 submissions from top research groups worldwide.

🏆 Our Results

  • 9th overall ranking in the competition
  • Highest detection accuracy among all participants: 0.78, surpassing the next best 0.67
  • Invited presentation at the NeurIPS 2025 AI-Face Workshop
  • Co-authorship invitation for the official NeurIPS 2025 competition summary paper

These results confirm ICI Innolabs’ state-of-the-art capabilities in designing fair, reliable, and high-performing AI systems that generalize across demographic groups — a key challenge for responsible AI.

🔬 Our Technical Approach

Our pipeline combined lightweight architectures for efficiency with large-scale backbones for robustness. It includes:

  • MobileNetV2-based models trained with adaptive augmentation pipelines (YCbCr, Gaussian noise)
  • A CLIP-L/14-based Effort detector used for the final submission ensemble
  • Integrated AIM experiment tracking, ensuring reproducibility and transparent metrics

We’re releasing our full training and inference pipeline on our github repo.

Go back