Benchmarks

Face Recognition

Face recognition accuracy (NIST Standards)

The table below shows Face SDK quality metrics based on the National Institute of Standards and Technology (NIST) Standards:

NIST Face Recognition Vendor Test (FRVT) 1:1	Score
VISA* True Acceptance Rate (@FAR 1E-6)	99.70%
Mugshot* True Acceptance Rate (@FAR 1E-5)	99.76%
Border* True Acceptance Rate (@FAR: 1E-6)	99.75%
WILD* True Acceptance Rate (@FAR 1E-4)	96.94%

Click here to see the description of image types from the table*

• VISA – Full Frontal image type. The images have a size of 252x300 pixels. The mean interocular distance (IOD) is 69 pixels.
• Mugshot – Full Frontal image type. The images are of variable sizes. The mean IOD is 113 pixels.
• Border – The images are taken with a camera, which is oriented by an attendant toward a cooperating subject. This is done under time constraints so there are role, pitch and yaw angle variations. Background illumination is sometimes strong, so the face is under-exposed. There is some perspective distortion due to close range images. Some faces are partially cropped.
• Wild – The images include many photojournalism-style images. Resolution varies very widely. The images are very unconstrained with wide yaw and pitch pose variation. Faces can be occluded, including hair and hands.

Face recognition accuracy on extended LFW dataset with Wild images

For this test we used extended LFW dataset (LFW dataset + our internal dataset). The set of mismatched pairs was increased and LFW errors were fixed to get accurate measurements at low FAR.

FAR	9.300 TAR (%)	12.30 TAR (%)	12.50 TAR (%)	12.100 TAR (%)	12.1000 TAR (%)
1e-4	98.7	98.9	99.2	99.4	99.5
1e-5	98.0	98.1	98.7	99.1	99.4
1e-6	96.6	96.7	97.9	98.5	99.3
1e-7	93.2	94.3	96.3	97.5	98.9
1e-8	32.2	88.0	89.5	94.1	97.7
1e-9	10.5	82.0	80.9	84.9	93.9

ROC-curve

Template extraction speed using CPU and GPU

Desktop

Method	GPU (NVIDIA GTX 1070)	CPU (Core i5-9400 4.0GHz)
12v1000	47 ms	442 ms
9v300	10 ms	292 ms
12v100	8 ms	49 ms
12v50	6 ms	21 ms
12v30	5 ms	12 ms

Mobile

Method	CPU (Qualcomm Snapdragon 845)
12v1000	7968ms
9v300	1960ms
12v100	801ms
12v50	270ms
12v30	150ms

Note: The speed test was performed using Google Pixel 3.

Face recognition speed for Core i7 4.5 GHz*

Recognition method	Template generation (ms)	Accelerated Matching 1:N (ms)			Matching 1:1 (ms)
		N = 104	N = 106	N = 107
6.7	40 (45**)	0.25	12.1	126	0.04
7.7	170 (180**)	0.25	12.1	126	0.04
8.7	20 (20**)	0.25	12.1	126	0.04
9.30	30	0.18	12.0	117	0.04
9.300	260 (125***)	0.18	12.0	117	0.04
9.1000	730 (305***)	0.18	12.0	117	0.04
9.30mask	20	0.18	12.0	117	0.04
9.300mask	160 (79***)	0.18	12.0	117	0.04
9.1000mask	290 (144***)	0.18	12.0	117	0.04
10.30	24 (16***)	0.18	12.0	117	0.04
10.100	40 (24***)	0.18	12.0	117	0.04
10.1000	690 (355***)	0.18	12.0	117	0.04
11.1000	865 (425***)	0.22	15.0	151	0.04
12.30	10***	0.20	15.0	120	0.04
12.50	18***	0.20	15.0	120	0.04
12.100	41***	0.20	15.0	120	0.04
12.1000	412***	0.40	20.0	170	0.04

* – characteristics specified in this table are given for a single-core CPU.
** – template extraction time when processing_less_memory_consumption was set to true in the FacerecService.createRecognizer call for creating a Recognizer.
*** – template extraction time using the AVX2 instruction set (see Face Recognition).

Note:

• Accelerated search time is given for k=1. As for larger values of k, the time will increase up to the search time without acceleration.
• Accelerated search is implemented only for the recognition methods 6.5, 6.6, 6.7, 7.3, 7.6, 7.7, 8.6, 8.7, 9.30, 9.300, 9.1000, 10v30, 10v100, 10v1000, 11v1000, 12.30, 12.50 , 12.100, 12.1000.
• To achieve this speed, the templates in the index must be located in order of creation (by using the Recognizer.processing or Recognizer.loadTemplate method).
• To achieve higher speed, use GPU (see GPU Usage).

Memory Characteristics

Recognition method	Serialized template size (Bytes)	Template size in RAM (Bytes)	Memory consumption* (MB)
6.7	536	636	105 (85**)
7.7	536	636	195 (163**)
8.7	536	636	52 (40**)
9.30	280	380	155
9.300	280	380	210
9.1000	280	380	290
10.30	280	380	160
10.100	280	380	180
10.1000	280	380	270
11.1000	296	396	480
12.30	230	330	160
12.50	261	361	170
12.100	337	437	180
12.1000	400	500	440

* – the amount of memory consumed doesn't depend on the number of the Recognizer objects created by this method
** – memory consumption when processing_less_memory_consumption was set to true in the FacerecService.createRecognizer call for Recognizer creation

Face Detection

Face Detection speed

Core i7 4.5 GHz (Single-Core)

Configuration file	Capture time (ms)
	640x480, 1 face	640x480, 4 faces	1280x720, 1 face	1280x720, 4 faces	1920x1080, 1 face	1920x1080, 4 faces
common_capturer4_fda.xml	13	25	34	49	81	103
common_capturer4_fda_with_angles.xml	282	387	260	356	273	370
common_capturer4_mesh.xml	18	47	39	72	87	735
common_capturer4_mesh_with_angles.xml	291	415	268	383	281	398
common_capturer_blf_fda_auto.xml	6-30	12-36	8-32	14-38	19-44	26-51
common_capturer_blf_fda_back.xml	30	36	32	38	44	51
common_capturer_blf_fda_front.xml	6	12	8	14	19	26
common_capturer_refa_fda_a.xml	644	650	512	518	580	586
common_capturer_uld_fda.xml (min_size=150)	12	18	13	19	21	28
common_capturer_uld_fda.xml (min_size=90)	58	70	60	73	77	91
common_capturer_uld_fda.xml (min_size=50)	253	272	253	273	281	302
common_capturer4_fda_singleface.xml	16	-	51	-	123	-
common_capturer4_mesh_singleface.xml	23	-	58	-	129	-

GPU

Configuration file	Capture time (ms)
	640x480, 1 face	640x480, 4 faces	1280x720, 1 face	1280x720, 4 faces	1920x1080, 1 face	1920x1080, 4 faces
common_capturer_blf_fda_auto.xml	4-5	10-12	6-8	13-14	17-20	24-27
common_capturer_blf_fda_back.xml	5	12	8	14	20	27
common_capturer_blf_fda_front.xml	4	10	6	13	17	24
common_capturer_refa_fda_a.xml	236	240	229	235	170	176
common_capturer_uld_fda.xml (min_size=150)	4	10	5	11	13	20
common_capturer_uld_fda.xml (min_size=90)	14	21	17	23	26	34
common_capturer_uld_fda.xml (min_size=50)	27	34	27	35	47	49

Note: Actual capture time may vary depending on the image content.

Click here to see the list of the capturer configuration files

File	Detector	Set of points	Angles (roll/yaw/pitch)	Description and use case
common_capturer4_fda.xml	lbf	fda	[-30;30][-60;60][-60;60]	Frontal face detector.
common_capturer4_fda_with_angles.xml	lbf	fda	[-90;90][-60;60][-60;60]	Frontal face detector. Adapted for a wide range of head rotation angles.
common_capturer4_fda_with_angles_noise.xml	lbf	fda	[-90;90][-60;60][-60;60]	Frontal face detector. Adapted for a wide range of head rotation angles. Suitable for images with high noise level.
common_capturer4_fda_singleface.xml	lbf	fda	[-30;30][-60;60][-60;60]	Only one frontal face is detected.
common_capturer4_fda_singleface_with_angles.xml	lbf	fda	[-90;90][-75;75][-60;60]	Only one frontal face is detected. The detector is adapted for a wide range of head rotation angles.
common_capturer4_fda_singleface_with_angles_noise.xml	lbf	fda	[-90;90][-75;75][-60;60]	Only one frontal face is detected. The detector is adapted for a wide range of head rotation angles. Suitable for images with high noise level.
common_capturer4_lbf.xml	lbf	doublelbf	[-30;30][-60;60][-60;60]	Frontal face detector.
common_capturer4_lbf_singleface.xml	lbf	doublelbf	[-30;30][-60;60][-60;60]	Only one frontal face is detected.
common_capturer4_mesh.xml	lbf	mesh	[-30;30][-60;60][-60;60]	Frontal face detector that allows you to get a 3D face mask.
common_capturer4_mesh_with_angles.xml	lbf	mesh	[-90;90][-60;60][-60;60]	Frontal face detector adapted for a wide range of head rotation angles that allows you to get a 3D face mask.
common_capturer4_mesh_with_angles_noise.xml	lbf	mesh	[-90;90][-60;60][-60;60]	Frontal face detector adapted for a wide range of head rotation angles and suitable for images with high noise level. The detector allows you to get a 3D face mask.
common_capturer4_mesh_singleface.xml	lbf	mesh	[-30;30][-60;60][-60;60]	Only one frontal face is detected. The detector allows you to get a 3D face mask.
common_capturer4_mesh_singleface_with_angles.xml	lbf	mesh	[-90;90][-75;75][-60;60]	Only one frontal face is detected. The detector is adapted for a wide range of head rotation angles and allows you to get a 3D face mask.
common_capturer4_mesh_singleface_with_angles_noise.xml	lbf	mesh	[-90;90][-75;75][-60;60]	Only one frontal face is detected. The detector is adapted for a wide range of head rotation angles, suitable for images with high noise level and allows you to get a 3D face mask.
common_capturer_blf_fda_front.xml	blf	fda	[-70;70][-90;90][-70;70]	Detection of large face images (the face should take up most of the frame size). Suitable for detection of masked faces.
common_capturer_blf_fda_back.xml	blf	fda	[-70;70][-90;90][-70;70]	Detection of several faces or small face images. Suitable for detection of masked faces.
common_capturer_blf_fda_auto.xml	blf	fda	[-70;70][-90;90][-70;70]	Detection of large and small face images (the parameters `resolution_width` and `min_face_size` should be specified in the configuration file). Suitable for detection of masked faces.
common_capturer_refa_fda_a.xml	refa	fda	[-70;70][-90;90][-70;70]	Face detector recommended for use in expert systems. The detector provides face detection with the largest coverage of rotation angles and maximum quality (including masked faces).
common_capturer_uld_fda.xml.xml	uld	fda	[-70;70][-90;90][-70;70]	Detection of large and small face images. Suitable for detection of masked faces.
common_video_capturer_fda.xml	lbf	fda	[-30;30][-60;60][-60;60]	Frontal face video tracker (RGB only).
common_video_capturer_lbf.xml	lbf	singlelbf	[-30;30][-60;60][-60;60]	Frontal face video tracker (RGB only).
common_video_capturer_mesh.xml	lbf	mesh	[-30;30][-60;60][-60;60]	Frontal face video tracker (RGB only) that allows you to get a 3D face mask.
fda_tracker_capturer.xml	lbf	fda	[-30;30][-60;60][-60;60]	Frontal face video tracker.
fda_tracker_capturer.w.xml	lbf	fda	[-30;30][-60;60][-60;60]	Frontal face video tracker that can be used in case of insufficient lighting. Note that false detections can occur a bit more often in this case.
fda_tracker_capturer_mesh.xml	lbf	fda	[-30;30][-60;60][-60;60]	Frontal face video tracker that allows you to get a 3D face mask.
fda_tracker_capturer_fake_detector.xml	lbf	fda	[-30;30][-60;60][-60;60]	Detection speed is higher because only fitter is used (no detector). Suitable only if a face takes up most of the image size.
fda_tracker_capturer_blf.xml	blf	fda	[-30;30][-60;60][-60;60]	Frontal face video tracker. Suitable for detection of masked faces.
fda_tracker_capturer_refa_a.xml	refa	fda	[-70;70][-90;90][-70;70]	Frontal face video tracker. Recommended for use in expert systems. The tracker provides face detection with the largest coverage of rotation angles and maximum quality (including masked faces).
fda_tracker_capturer_uld_fda.xml.xml	uld	fda	[-70;70][-90;90][-70;70]	Frontal face video tracker that can be used to detect faces of different size. Suitable for detection of masked faces.
manual_capturer_fda.xml	lbf	fda	[-30;30][-60;60][-60;60]	Eye points should be manually specified. The remaining points are calculated based on the eye points.
manual_capturer_mesh.xml	lbf	mesh	[-30;30][-60;60][-60;60]	Eye points should be manually specified. The remaining points are calculated based on the eye points that allows you to get a 3D face mask.
video_worker_fdatracker_refa_fda.xml	refa	fda	[-70;70][-90;90][-70;70]	Face detector recommended for use in expert systems. The detector provides face detection with the largest coverage of rotation angles and maximum quality (including masked faces).

Face detection speed using CPU and GPU

The tables in this section show the speed measurements for Face SDK modules using CPU and GPU.

Detector	GPU (NVIDIA GTX 1080 Ti)	CPU (Core i7-7700K 4.4GHz)
ULD	35ms	200ms
BLF	14ms	38ms
REFA	235ms	518ms

Speed is indicated for 4 faces in an image with a resolution of 1280x720.

Estimation of Face Attributes

Estimation accuracy

Face Attributes	Accuracy
Gender	95%
Emotion	80%
Age	3.95 y/o (mean Average Error)
No mask (face without a mask)	99%
Has mask (face with a mask)	97%