STMicroelectronics/ssdlite_mobilenetv3large_pt

SSDLite MobileNetV3Large

Use case : Object detection

Model description

SSDLite MobileNetV3Large is a high-capacity lightweight single-shot object detection model optimized for real-time inference on mobile and edge devices while providing improved accuracy compared to standard SSDLite MobileNetV2.

It combines the SSDLite framework with MobileNetV3Large as the backbone. MobileNetV3Large uses inverted residual blocks with expanded width, providing stronger representational power while maintaining the efficiency benefits of MobileNetV3.
The SSDLite detection head predicts object locations and class probabilities in a single forward pass, making the model suitable for real-time detection on resource-constrained platforms, especially when higher accuracy is needed.

The ssdlite_mobilenetv3large_pt variant is implemented in PyTorch and is used in applications where low latency, reasonable memory footprint, and better accuracy are desired for edge and mobile deployments.

Network information

Network information	Value
Framework	Torch
Quantization	Int8
Provenance	torchvision GitHub
Paper	SSDLite MobileNetV3

The model is quantized to int8 using ONNX Runtime and exported for efficient deployment.

Network inputs / outputs

For an image resolution of NxM and NC classes

Input Shape	Description
(1, W, H, 3)	Single NxM RGB image with UINT8 values between 0 and 255

Output Shape	Description
(1, 3000,(1+NC) and (1,3000,4))	Model returns two output vectors of bounding boxes where first output returns confidence for each class (+ background class) and second output returns bounding box coordinates (x1, y1, x2,y2)

Recommended Platforms

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

Performances

Metrics

Measures are done with default STEdgeAI Core configuration with enabled input / output allocated option.

Reference NPU memory footprint based on COCO dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Series	Internal RAM (KiB)	External RAM (KiB)	Weights Flash (KiB)	STEdgeAI Core version
ssdlite_mobilenetv3large_pt	COCO	Int8	300x300x3	STM32N6	2484.27	0	3592.83	3.0.0

Reference NPU inference time based on COCO dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Board	Execution Engine	Inference time (ms)	Inf / sec	STEdgeAI Core version
ssdlite_mobilenetv3large_pt	COCO	Int8	300x300x3	STM32N6570-DK	NPU/MCU	34.62	28.89	3.0.0

Reference NPU memory footprint based on COCO Person dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Series	Internal RAM (KiB)	External RAM (KiB)	Weights Flash (KiB)	STEdgeAI Core version
ssdlite_mobilenetv3large_pt	COCO-Person	Int8	300x300x3	STM32N6	2247.37	0	2592.98	3.0.0

Reference NPU inference time based on COCO Person dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Board	Execution Engine	Inference time (ms)	Inf / sec	STEdgeAI Core version
ssdlite_mobilenetv3large_pt	COCO-Person	Int8	300x300x3	STM32N6570-DK	NPU/MCU	31.45	31.80	3.0.0

Reference NPU memory footprint based on VOC dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Series	Internal RAM (KiB)	External RAM (KiB)	Weights Flash (KiB)	STEdgeAI Core version
ssdlite_mobilenetv3large_pt	VOC	Int8	300x300x3	STM32N6	2242.98	0	2833.48	3.0.0

Reference NPU inference time based on VOC dataset (see Accuracy for details on dataset)

Model	Dataset	Format	Resolution	Board	Execution Engine	Inference time (ms)	Inf / sec	STEdgeAI Core version
ssdlite_mobilenetv3large_pt	VOC	Int8	300x300x3	STM32N6570-DK	NPU/MCU	32.16	31.09	3.0.0

AP on COCO dataset

Dataset details: link , License CC BY 4.0, Number of classes: 80

Model	Format	Resolution	AP50
ssdlite_mobilenetv3large_pt	Float	3x300x300	27.72
ssdlite_mobilenetv3large_pt	Int8	3x300x300	26.78

* EVAL_IOU = 0.5, NMS_THRESH = 0.5, SCORE_THRESH = 0.001, MAX_DETECTIONS = 100

AP on COCO-Person dataset

Dataset details: link , License CC BY 4.0 , Number of classes: 1

Model	Format	Resolution	AP50
ssdlite_mobilenetv3large_pt	Float	3x300x300	39.24
ssdlite_mobilenetv3large_pt	Int8	3x300x300	38.20

* EVAL_IOU = 0.5, NMS_THRESH = 0.5, SCORE_THRESH = 0.001, MAX_DETECTIONS = 100

AP on VOC dataset

Dataset details: link , License , Number of classes: 20

Model	Format	Resolution	AP50
ssdlite_mobilenetv3large_pt	Float	3x300x300	65.84
ssdlite_mobilenetv3large_pt	Int8	3x300x300	65.49

* EVAL_IOU = 0.5, NMS_THRESH = 0.5, SCORE_THRESH = 0.001, MAX_DETECTIONS = 100

Retraining and Integration in a simple example:

Please refer to the stm32ai-modelzoo-services GitHub here

References

• COCO
  Lin et al., Microsoft COCO: Common Objects in Context

• PASCAL VOC
  Everingham et al., The PASCAL Visual Object Classes (VOC) Challenge

• SSDLite
  Howard et al., MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications (SSDLite section)

• MobileNetV3
  Howard et al., Searching for MobileNetV3