How Accurate is RetinaFace?
RetinaFace is one of the most accurate face detectors ever benchmarked. On the WIDER FACE dataset, RetinaFace achieves Average Precision (AP) scores that surpass earlier state-of-the-art methods by significant margins, particularly on the hard subset which contains tiny, occluded, and heavily distorted faces.
WIDER FACE Benchmark Results
RetinaFace (ResNet-50) AP Scores on WIDER FACE Validation
- Easy: ~96.9% AP
- Medium: ~96.1% AP
- Hard: ~91.4% AP
These numbers represent a substantial improvement over prior methods. Earlier detectors like MTCNN achieve approximately 85% AP on the hard set, meaning RetinaFace misses significantly fewer challenging faces.
Face Recognition Downstream Accuracy
Studies that replaced earlier detection stages with RetinaFace in face recognition pipelines reported consistent improvements in verification accuracy on datasets like LFW, IJB-B, and IJB-C, demonstrating that the benefits of improved detection flow through to the end task.
Conditions That Reduce Accuracy
- Extreme out-of-plane rotation (beyond 80 degrees from frontal)
- Heavy occlusion covering more than 60% of the face
- Very low resolution faces (below 10x10 pixels)
- Severe motion blur or extreme compression artifacts
- Unusual lighting such as strong infrared or silhouette conditions
Practical Accuracy in Production
In controlled production environments with good lighting and frontal faces, RetinaFace effective precision and recall rates typically exceed 99% for face detection. Real-world accuracy in typical business applications tends to be higher than benchmark scores suggest.
Conclusion
RetinaFace is one of the most accurate single-stage face detectors available, with benchmark results that demonstrate robustness across a wide range of challenging conditions.
