BlazeFace Front 128x128 Quantized

Use case : `Object detection`

Model description

BlazeFace Front 128x128 is a lightweight and efficient face detection model optimized for real-time applications on embedded devices. It is a variant of the BlazeFace architecture, designed specifically for detecting frontal faces at a resolution of 128x128 pixels. The model is quantized to int8 format using TensorFlow Lite converter to reduce memory footprint and improve inference speed on resource-constrained hardware.

BlazeFace is known for its fast inference and accuracy, making it suitable for applications such as face tracking, augmented reality, and user authentication.

Network information

Network information	Value
Framework	TensorFlow Lite
Quantization	int8
Input resolution	128x128
Provenance	https://github.com/PINTO0309/PINTO_model_zoo/tree/main/030_BlazeFace

Network inputs / outputs

Input Shape	Description
(1, 128, 128, 3)	Single 128x128 RGB image with FLOAT32 values between -1 and 1

Output Shape	Description
(1, 512, 16)	FLOAT32 tensor containing bounding box coordinates and keypointss for detected faces
(1, 512, 1)	FLOAT32 tensor containing confidence scores for detected faces
(1, 384, 16)	FLOAT32 tensor containing bounding box coordinates and keypointss for detected faces
(1, 384, 1)	FLOAT32 tensor containing confidence scores for detected faces

Recommended Platforms

Platform	Supported	Recommended
STM32L0	[]	[]
STM32L4	[]	[]
STM32U5	[]	[]
STM32H7	[]	[]
STM32MP1	[]	[]
STM32MP2	[]	[]
STM32N6	[x]	[x]

Performances

Metrics

Performance metrics are measured using default STM32Cube.AI configurations with input/output allocated buffers.

Model	Dataset	Format	Resolution	Series	Internal RAM (KB)	External RAM (KB)	Weights Flash (KB)	STM32Cube.AI version	STEdgeAI Core version
BlazeFace Front 128x128 per channel	WIDER FACE (frontal)	Int8	128x128x3	STM32N6	528	0	150.97	10.2.0	2.2.0

Reference NPU inference time (example)

Model	Dataset	Format	Resolution	Board	Execution Engine	Inference time (ms)	Inf / sec	STM32Cube.AI version	STEdgeAI Core version
BlazeFace Front 128x128 per channel	WIDER FACE (frontal)	Int8	128x128x3	STM32N6570-DK	NPU/MCU	5.09	196.3	10.2.0	2.2.0

Integration and support

For integration examples and additional services, please refer to the STM32 AI model zoo services repository:
https://github.com/STMicroelectronics/stm32ai-modelzoo-services

References

BlazeFace paper: https://arxiv.org/abs/1907.05047
MediaPipe BlazeFace model repository: https://github.com/PINTO0309/PINTO_model_zoo/tree/main/030_BlazeFace
WIDER FACE dataset: http://shuoyang1213.me/WIDERFACE/