Research map / Image generation / Autoregressive and latent generation
CLIB-FIQA: Face Image Quality Assessment with Confidence Calibration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CLIB-FIQA, a CLIP-based face image quality assessment method that jointly learns quality factors (blur, pose, expression, occlusion, lighting) via text prompts and calibrates quality-anchor confidence using a merged-factor distribution.
Problem
Quality-fitting-based face image quality assessment (FIQA) relies on recognition-model quality anchors that are treated as equally confident, causing a fitting bottleneck from inaccurate anchors.
Approach
Adapt CLIP with textual descriptions of multiple quality factors to jointly learn a joint distribution of factors and quality levels, then compute a confidence from the JS divergence between the merged-factor distribution and the predicted quality distribution, and use it to scale the soft-mapping shape parameter for confidence-calibrated quality fitting.
Key results
- Under ArcFace, CLIB-FIQA reduces average pAUC by 1.81% and average AUC by 7.6% versus the best competitor (per the paper's text).
- Under CosFace, average AUC is reduced by about 3.34% relative to CR-FIQA, while pAUC is not best on LFW.
- Under AdaFace, the paper reports the method has better average pAUC/AUC than the compared methods; table shows CLIB-FIQA avg pAUC 0.664 vs CR-FIQA 0.683 and avg AUC 0.412 vs CR-FIQA 0.451.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLIB-FIQA instead of CR-FIQA | FIQA with ArcFace deployed for recognition · Eight benchmarks | pAUC@FMR=1E−3 (Avg.): 0.660 -> 0.648 | Helped |
| Use CLIB-FIQA instead of CR-FIQA | FIQA with ArcFace deployed for recognition · Eight benchmarks | AUC@FMR=1E−3 (Avg.): 0.431 -> 0.398 | Helped |
| Use CLIB-FIQA instead of CR-FIQA | FIQA with CosFace deployed for recognition · Eight benchmarks | pAUC@FMR=1E−3 (Avg.): 0.688 -> 0.665 | Helped |
| Use CLIB-FIQA instead of CR-FIQA | FIQA with CosFace deployed for recognition · Eight benchmarks | AUC@FMR=1E−3 (Avg.): 0.439 -> 0.419 | Helped |
| Use CLIB-FIQA instead of CR-FIQA | FIQA with AdaFace deployed for recognition · Eight benchmarks | pAUC@FMR=1E−3 (Avg.): 0.683 -> 0.664 | Helped |
| Use CLIB-FIQA instead of CR-FIQA | FIQA with AdaFace deployed for recognition · Eight benchmarks | AUC@FMR=1E−3 (Avg.): 0.451 -> 0.412 | Helped |
| Add quality-factor learning through the merged-factor distribution (ablation) | CLIP FIQA on AdaFace · Eight benchmarks | pAUC@FMR=1E−3 (Avg.): 0.705 -> 0.763 | Hurt |
| Add all five quality factors and confidence calibration (ablation) | CLIP FIQA on AdaFace · Eight benchmarks | pAUC@FMR=1E−3 (Avg.): 0.669 -> 0.654 | Helped |
Surprising findings
- The paper's own tables show CLIB-FIQA does not win on every benchmark; e.g., under ArcFace it has higher pAUC on LFW than some baselines and loses on LFW AUC to FaceQnet-like baselines in some settings.
Limitations
- Relies on automatically generated quality-factor labels from heuristic classifiers (CPBD, yaw angle, WiderFace-trained CNNs) for MS1MV2.
- Still depends on a frozen face recognition model to produce quality anchors.
- Some benchmarks (e.g., LFW pAUC under AdaFace, CFP-FP under ArcFace) show weaker results than baselines.
Experimental setup
Datasets: MS1MV2 WiderFace
Benchmarks: LFW CFP-FP CPLFW XQLFW CALFW AgeDB Adience TinyFace
Metrics: AUC pAUC EVRC FNMR
Baselines: PFE SER-FIQ MagFace FaceQAN FaceQnet PCNet SDD-FIQA CR-FIQA
Builds on: CLIP CR-FIQA PCNet SDD-FIQA FaceQnet
Closest papers by meaning
- DSL-FIQA: Assessing Facial Image Quality via Dual-Set Degradation Learning and Landmark-Guided Transformer (CVPR 2024 · similarity 0.88)
- Integrating Global Context Contrast and Local Sensitivity for Blind Image Quality Assessment (ICML 2024 · similarity 0.83)
- RepFace: Refining Closed-Set Noise with Progressive Label Correction for Face Recognition (AAAI 2025 · similarity 0.83)
- Teaching Large Language Models to Regress Accurate Image Quality Scores Using Score Distribution (CVPR 2025 · similarity 0.82)
- FaceBench: A Multi-View Multi-Level Facial Attribute VQA Dataset for Benchmarking Face Perception MLLMs (CVPR 2025 · similarity 0.82)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.