Benchmarks

Trueface SDK Benchmarks

A typical 1 to N face recognition pipeline involves the following steps:

Preprocess Image
Face and landmark detection
Face recognition template extraction
1 to N identification search

To estimate the total pipeline time, add up the time for each individual step. You may also need to account for the time taken to decode your video stream.

If you chose to run face recognition on every face in your image, you will need repeat steps 3 and 4 for every detected face.

Trueface SDK Benchmarks: CPU

Operation	Ram Req.	i9-11900H	AWS c6i.2xlarge (x86)
Preprocess image (JPG from disk)	-	4.6 ms	6.4 ms
Preprocess image (encoded JPG in memory)	-	4.2 ms	5.3 ms
Preprocess image (RGB pixels array in memory)	-	0.08 ms	0.16 ms
Face image orientation detection	68 Mb	44.6 ms	48.0 ms
Face and landmark detection	42 Mb	7.9 ms	12.5 ms
106 face landmark detection	39 Mb	2.2 ms	2.7 ms
Head orientation detection (yaw, pitch, roll)	52 Mb	0.4 ms	0.5 ms
Face image blur detection	109 Mb	8.4 ms	12.2 ms
Face template quality estimation	115 Mb	2.5 ms	3.2 ms
Blink detection	100 Mb	6.1 ms	9.7 ms
Mask detection	118 Mb	3.0 ms	3.6 ms
Eyeglasses detection	40 Mb	0.3 ms	0.3 ms
Passive spoof detection	140 Mb	24.1 ms	30.7 ms
Object detection (fast)	70 Mb	29.3 ms	39.8 ms
Object Detection (accurate)	80 Mb	121.5 ms	155.2 ms
Face recognition LITE model	54 Mb	3.2 ms	3.4 ms
Face recognition LITE_V2 model	76 Mb	6.3 ms	10.2 ms
Face recognition LITE_V3 model	100 Mb	6.5 ms	8.5 ms
Face recognition TFV5_2 model	260 Mb	36.0 ms	49.9 ms
Face recognition TFV6 model	260 Mb	36.1 ms	49.9 ms
Face recognition TFV7 model	500 Mb	67.2 ms	94.2 ms
1 to N identification search (N = 1,000,000) TFV7	1250 bytes / template	26.9 ms	21.4 ms
1 to N batch identification search (N = 1,000,000) TFV7	1250 bytes / template	10.9 ms	15.5 ms

1 to N identification search times scale linearly relative to collection size (ex. for a collection of size 10,000, divide the above reported times by 100). frVectorCompression flag set to true for 1 to N benchmarks.

All benchmarks performed using 1280x720 pixel images containing 1 face or object with CPU only. Ram usage refers to maximum resident memory. Batch identification tested with 100 probe templates, and is used to increase throughput. Enrollment template size represents conservative average case, but it can be variable due to variable length in identity string. All benchmarks run on Ubuntu 20.04. A Licensee’s results may vary depending on the SDK model and the nature of its input images and data.

Trueface SDK Benchmarks: GPU

Operation	VRAM Usage	RTX 4090	AWS g4dn.xlarge (T4)
Preprocess image (JPG from disk)	-	6.3 ms	7.7 ms
Preprocess image (encoded JPG in memory)	-	6.3 ms	7.6 ms
Preprocess image (RGB pixels array in memory)	-	0.12 ms	0.14 ms
Face image orientation detection	506 Mb	0.61 ms	2.2 ms
Face image orientation detection, batch size = 16	506 Mb	0.47 ms	1.8 ms
Face and landmark detection	130 Mb	1.6 ms	1.8 ms
106 face landmark detection	148 Mb	0.36 ms	0.47 ms
106 face landmark detection, batch size = 16	148 Mb	0.08 ms	0.20 ms
Head orientation detection (yaw, pitch, roll)	160 Mb	0.54 ms	0.80 ms
Face image blur detection	254 Mb	0.27 ms	0.82 ms
Face image blur detection, batch size = 16	254 Mb	0.09 ms	0.41 ms
Face template quality estimation	250 Mb	0.23 ms	0.69 ms
Face template quality estimation, batch size = 16	250 Mb	0.08 ms	0.33 ms
Blink detection	154 Mb	0.65 ms	0.93 ms
Blink detection, batch size = 16	154 Mb	0.35 ms	0.55 ms
Mask detection	160 Mb	0.50 ms	0.9 ms
Mask detection, batch size = 16	160 Mb	0.07 ms	0.17 ms
Passive spoof detection	278 Mb	5.3 ms	6.9 ms
Object detection (fast)	160 Mb	4.2 ms	6.3 ms
Object Detection (accurate)	234 Mb	13.6 ms	21.9 ms
Face recognition LITE_V2 model	144 Mb	0.66 ms	0.95 ms
Face recognition LITE_V2 model, batch size = 16	144 Mb	0.12 ms	0.43 ms
Face recognition LITE_V3 model	214 Mb	0.55 ms	1.5 ms
Face recognition LITE_V3 model, batch size = 16	214 Mb	0.17 ms	0.77 ms
Face recognition TFV5_2 model	332 Mb	1.1 ms	3.1 ms
Face recognition TFV5_2 model, batch size = 16	332 Mb	0.36 ms	1.8 ms
Face recognition TFV6 model	328 Mb	1.2 ms	3.2 ms
Face recognition TFV6 model, batch size = 16	328 Mb	0.37 ms	1.8 ms
Face recognition TFV7 model	430 Mb	2.2 ms	6.0 ms
Face recognition TFV7 model, batch size = 16	430 Mb	0.69 ms	3.2 ms

1 to N identification search is performed on CPU only, so refer to CPU times.

How are batching times reported:

With a batch size of 4, we generate 4 face recognition templates at the same time. The total time taken to generate those templates is 16.4 ms, meaning the average time per template is 4.1 ms.

With a batch size of 64, we generate 64 face recognition templates at the same time. The total time taken to generate those templates is 136.96 ms, meaning the average time per template is 2.14 ms.

Operations which do not show a batch size only support a batch size of 1 at this time.

All benchmarks performed using 1280x720 pixel images containing 1 face or object with GPU enabled. GPU benchmarks use FP16 inference. A Licensee’s results may vary depending on the SDK model and the nature of its input images and data.

Trueface On-Prem Benchmarks

Operation	Ram Req.	AWS c6i.2xlarge (CPU)	AWS g4dn.xlarge (GPU)
Templatize Face	2.1 Gb	104 ms	31 ms
Match Faces	2.1 Gb	249 ms	60 ms
Enroll Face	2.1 Gb	93 ms	30 ms
Identify Face (N = 1000)	2.1 Gb	119 ms	31 ms
Spoof Detection	2.5 Gb	78 ms	35 ms

All benchmarks performed using 1280x720 px jpg image on disk using default smallest face height of 100 and TFV5_2 face recognition model. Requests were sent from the same machine running the PTOP server in order to avoid any network overhead in the measurements. A Licensee’s results may vary depending on the SDK model and the nature of its input images and data.