SCIENCE@home

RankCORE: Self-Supervised Ranking-Aware Correlation Optimized Regression for Face Image Quality Assessment

Abhishek Joshi,Aman Agarwal

Published 2025 in 2025 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)

ABSTRACT

Generic Face Image Quality Assessment (GFIQA) remains a core challenge in face analytics, where subtle degradations such as blur, misalignment, and occlusion can severely impact downstream task performance. Existing methods either rely on expensive human annotations or regress to pseudo-labels, often suffering from scalability issues and suboptimal generalization, particularly in ranking images with fine-grained degradations. In this paper, we introduce RankCORE, a unified and efficient framework for perceptual face quality estimation that addresses both learning biases and ranking inconsistencies. Our method builds on three key innovations: (1) a Self-Supervised Adaptive Ranking loss (SSAR) that dynamically modulates ranking margins based on sample difficulty, enabling precise discrimination of subtle degradations without the need for explicit quality labels; (2) a Score-Stratified Uniform Sampler (SSUS) that combats distribution imbalance by equalizing the gradient contribution across the quality spectrum; and (3) a novel CoReFace loss, which jointly optimizes local prediction accuracy and global rank correlation. Our lightweight model (0.95M parameters) achieves impressive scores in the ongoing FIQA 2025 Challenge benchmark with PLCC, SROCC metrics, while using just 19% of the parameter budget. Additionally, RankCORE outperforms state-of-the-art approaches on public benchmarks: CGFIQA-40k and GFIQA-20k, demonstrating strong generalization and practical deployment viability. Our contributions mark a significant step toward scalable, efficient, and ranking-consistent FIQA solutions. Interestingly, our method tops the FIQA leaderboard in terms of number of parameters used while maintaining competitive accuracy.

PUBLICATION RECORD

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-48 of 48 references · Page 1 of 1

CITED BY