README.md

InsightFace Model Zoo

🔔 ALL models are available for non-commercial research purposes only.

0. Python Package models

To check the detail of insightface python package, please see here.

To install: pip install -U insightface

To use the specific model pack:

model_pack_name = 'buffalo_l'
app = FaceAnalysis(name=model_pack_name)

Name in bold is the default model pack in latest version.

Name	Detection Model	Recognition Model	Alignment	Attributes	Model-Size
antelopev2	RetinaFace-10GF	ResNet100@Glint360K	2d106 & 3d68	Gender&Age	407MB
buffalo_l	RetinaFace-10GF	ResNet50@WebFace600K	2d106 & 3d68	Gender&Age	326MB
buffalo_m	RetinaFace-2.5GF	ResNet50@WebFace600K	2d106 & 3d68	Gender&Age	313MB
buffalo_s	RetinaFace-500MF	MBF@WebFace600K	2d106 & 3d68	Gender&Age	159MB
buffalo_sc	RetinaFace-500MF	MBF@WebFace600K	-	-	16MB

Recognition accuracy of python library model packs:

Name	MR-ALL	African	Caucasian	South Asian	East Asian	LFW	CFP-FP	AgeDB-30	IJB-C(E4)
buffalo_l	91.25	90.29	94.70	93.16	74.96	99.83	99.33	98.23	97.25
buffalo_s	71.87	69.45	80.45	73.39	51.03	99.70	98.00	96.58	95.02

buffalo_m has the same accuracy with buffalo_l.

buffalo_sc has the same accuracy with buffalo_s.

(Note that almost all ONNX models in our model_zoo can be called by python library.)

1. Face Recognition models.

Definition:

The default training loss is margin based softmax if not specified.

MFN: MobileFaceNet

MS1MV2: MS1M-ArcFace

MS1MV3: MS1M-RetinaFace

MS1M_MegaFace: MS1MV2+MegaFace_train

_pfc: using Partial FC, with sample-ratio=0.1

MegaFace: MegaFace identification test, with gallery=1e6.

IJBC: IJBC 1:1 test, under FAR<=1e-4.

BDrive: BaiduDrive

GDrive: GoogleDrive

List of models by MXNet and PaddlePaddle:

Backbone	Dataset	Method	LFW	CFP-FP	AgeDB-30	MegaFace	Link.
R100 (mxnet)	MS1MV2	ArcFace	99.77	98.27	98.28	98.47	BDrive, GDrive
MFN (mxnet)	MS1MV1	ArcFace	99.50	88.94	95.91	-	BDrive, GDrive
MFN (paddle)	MS1MV2	ArcFace	99.45	93.43	96.13	-	pretrained model, inference model
iResNet50 (paddle)	MS1MV2	ArcFace	99.73	97.43	97.88	-	pretrained model, inference model

List of models by various depth IResNet and training datasets:

Backbone	Dataset	MR-ALL	African	Caucasian	South Asian	East Asian	Link(onnx)
R100	Casia	42.735	39.666	53.933	47.807	21.572	GDrive
R100	MS1MV2	80.725	79.117	87.176	85.501	55.807	GDrive
R18	MS1MV3	68.326	62.613	75.125	70.213	43.859	GDrive
R34	MS1MV3	77.365	71.644	83.291	80.084	53.712	GDrive
R50	MS1MV3	80.533	75.488	86.115	84.305	57.352	GDrive
R100	MS1MV3	84.312	81.083	89.040	88.082	62.193	GDrive
R18	Glint360K	72.074	68.230	80.575	75.852	47.831	GDrive
R34	Glint360K	83.015	79.907	88.620	86.815	60.604	GDrive
R50	Glint360K	87.077	85.272	91.617	90.541	66.813	GDrive
R100	Glint360K	90.659	89.488	94.285	93.434	72.528	GDrive

List of models by IResNet-50 and different training datasets:

Dataset	MR-ALL	African	Caucasian	South Asian	East Asian	LFW	CFP-FP	AgeDB-30	IJB-C(E4)	Link(onnx)
CISIA	36.794	42.550	55.825	49.618	19.611	99.450	95.214	94.900	87.220	GDrive
CISIA_pfc	37.107	38.934	53.823	48.674	19.927	99.367	95.429	94.600	84.970	GDrive
VGG2	38.578	35.259	54.304	44.081	24.095	99.550	97.410	95.080	91.220	GDrive
VGG2_pfc	40.673	36.767	60.180	49.039	24.255	99.683	98.529	95.400	92.490	GDrive
GlintAsia	62.663	49.531	64.829	57.984	61.743	99.583	93.186	95.400	91.500	GDrive
GlintAsia_pfc	63.149	50.366	65.227	57.936	61.820	99.650	93.029	95.233	91.140	GDrive
MS1MV2	77.696	74.596	84.126	82.041	51.105	99.833	98.083	98.083	96.140	GDrive
MS1MV2_pfc	77.738	74.728	84.883	82.798	52.507	99.783	98.071	98.017	96.080	GDrive
MS1M_MegaFace	78.372	74.138	82.251	77.223	60.203	99.750	97.557	97.400	95.350	GDrive
MS1M_MegaFace_pfc	78.773	73.690	82.947	78.793	57.566	99.800	97.870	97.733	95.400	GDrive
MS1MV3	82.522	77.172	87.028	86.006	60.625	99.800	98.529	98.267	96.580	GDrive
MS1MV3_pfc	81.683	78.126	87.286	85.542	58.925	99.800	98.443	98.167	96.430	GDrive
Glint360k	86.789	84.749	91.414	90.088	66.168	99.817	99.143	98.450	97.130	GDrive
Glint360k_pfc	87.077	85.272	91.616	90.541	66.813	99.817	99.143	98.450	97.020	GDrive
WebFace600K	90.566	89.355	94.177	92.358	73.852	99.800	99.200	98.100	97.120	GDrive
WebFace600K_pfc	89.951	89.301	94.016	92.381	73.007	99.817	99.143	98.117	97.010	GDrive
Average	69.247	65.908	77.121	72.819	52.014	99.706	97.374	96.962	93.925
Average_pfc	69.519	65.898	77.497	73.213	51.853	99.715	97.457	96.965	93.818

List of models by MobileFaceNet and different training datasets:

FLOPS: 450M FLOPs

Model-Size: 13MB

Dataset	MR-ALL	African	Caucasian	South Asian	East Asian	LFW	CFP-FP	AgeDB-30	IJB-C(E4)	Link(onnx)
WebFace600K	71.865	69.449	80.454	73.394	51.026	99.70	98.00	96.58	95.02	-

2. Face Detection models.

2.1 RetinaFace

In RetinaFace, mAP was evaluated with multi-scale testing.

m025: means MobileNet-0.25

Impelmentation	Easy-Set	Medium-Set	Hard-Set	Link
RetinaFace-R50	96.5	95.6	90.4	BDrive, GDrive
RetinaFace-m025(yangfly)	-	-	82.5	BDrive(nzof), GDrive
BlazeFace-FPN-SSH (paddle)	91.9	89.8	81.7%	pretrained model, inference model

2.2 SCRFD

In SCRFD, mAP was evaluated with single scale testing, VGA resolution.

2.5G: means the model cost 2.5G FLOPs while the input image is in VGA(640x480) resolution.

_KPS: means this model can detect five facial keypoints.

Name	Easy	Medium	Hard	FLOPs	Params(M)	Infer(ms)	Link(pth)
SCRFD_500M	90.57	88.12	68.51	500M	0.57	3.6	GDrive
SCRFD_1G	92.38	90.57	74.80	1G	0.64	4.1	GDrive
SCRFD_2.5G	93.78	92.16	77.87	2.5G	0.67	4.2	GDrive
SCRFD_10G	95.16	93.87	83.05	10G	3.86	4.9	GDrive
SCRFD_34G	96.06	94.92	85.29	34G	9.80	11.7	GDrive
SCRFD_500M_KPS	90.97	88.44	69.49	500M	0.57	3.6	GDrive
SCRFD_2.5G_KPS	93.80	92.02	77.13	2.5G	0.82	4.3	GDrive
SCRFD_10G_KPS	95.40	94.01	82.80	10G	4.23	5.0	GDrive

3. Face Alignment models.

2.1 2D Face Alignment

Impelmentation	Points	Backbone	Params(M)	Link(onnx)
Coordinate-regression	106	MobileNet-0.5	1.2	GDrive

2.2 3D Face Alignment

Impelmentation	Points	Backbone	Params(M)	Link(onnx)
-	68	ResNet-50	34.2	GDrive

2.3 Dense Face Alignment

4. Face Attribute models.

4.1 Gender&Age

Training-Set	Backbone	Params(M)	Link(onnx)
CelebA	MobileNet-0.25	0.3	GDrive

4.2 Expression