RuteNL/SCRFD-face-detection-ONNX

SCRFD face detection models converted to ONNX.

Pretrained models retrieved from https://github.com/deepinsight/insightface/tree/master/detection/scrfd#pretrained-models

Face detection models

This crate uses SCRDF face detection models. The following models are available:

The naming convention for the ONNX models indicates the computational complexity (measured in FLOPs) and whether the model includes 5 facial keypoints predictions in addition to standard bounding boxes.

Name	Easy	Medium	Hard	FLOPs	Params(M)	Infer(ms)	BBox	Facial Keypoints
500m.onnx	90.57	88.12	68.51	500M	0.57	3.6	✅	❌
1g.onnx	92.38	90.57	74.80	1G	0.64	4.1	✅	❌
34g.onnx	96.06	94.92	85.29	34G	9.80	11.7	✅	❌
2.5g_bnkps.onnx	93.80	92.02	77.13	2.5G	0.82	4.3	✅	✅
10g_bnkps.onnx	95.40	94.01	82.80	10G	4.23	5.0	✅	✅
34g_gnkps.onnx	?	?	?	34G	?	?	✅	✅

Keypoints (kps) and Normalization Types (bn vs gn)

• kps: Denotes models that output 5 facial landmarks (keypoints) in addition to the standard bounding boxes.
• bnkps: Models trained using Batch Normalization (BN). These often have lower false-positive rates and high recall on general datasets. However, they occasionally struggle with producing accurate landmarks for faces that are rotated past 90 degrees or are unusually large.
• gnkps: Models trained using Group Normalization (GN). These variants (e.g., 34g_gnkps or 10g_gnkps) were explicitly developed to fix issues with very large faces that the bnkps models exhibited. While they improve landmark quality on large or rotated faces, they might have slightly lower general recall than bnkps.