ST YOLO LC V1 quantized

Use case : `Object detection`

Model description

ST Yolo LC v1 is a real-time object detection model targeted for real-time processing implemented in Tensorflow.

The model is quantized in int8 format using tensorflow lite converter.

Network information

Network information	Value
Framework	TensorFlow Lite
Quantization	int8
Paper	https://pjreddie.com/media/files/papers/YOLO9000.pdf

The models are quantized using tensorflow lite converter.

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, WxH, NAx(5+NC))	FLOAT values Where WXH is the resolution of the output grid cell, NA is the number of anchors and NC is the number of classes

Recommended Platforms

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

Performances

Metrics

Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.

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

Model	Dataset	Format	Resolution	Series	Internal RAM (KiB)	Weights Flash (KiB)	STM32Cube.AI version	STEdgeAI Core version
st_yolo_lc_v1	COCO-Person	Int8	192x192x3	STM32N6	252	316.69	10.2.0	2.2.0
st_yolo_lc_v1	COCO-Person	Int8	224x224x3	STM32N6	343	316.69	10.2.0	2.2.0
st_yolo_lc_v1	COCO-Person	Int8	256x256x3	STM32N6	576	316.69	10.2.0	2.2.0

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	STM32Cube.AI version	STEdgeAI Core version
st_yolo_lc_v1	COCO-Person	Int8	192x192x3	STM32N6570-DK	NPU/MCU	1.96	510.2	10.2.0	2.2.0
st_yolo_lc_v1	COCO-Person	Int8	224x224x3	STM32N6570-DK	NPU/MCU	2.36	423.73	10.2.0	2.2.0
st_yolo_lc_v1	COCO-Person	Int8	256x256x3	STM32N6570-DK	NPU/MCU	3.02	331.13	10.2.0	2.2.0

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

Model	Format	Resolution	Series	Activation RAM	Runtime RAM	Weights Flash	Code Flash	Total RAM	Total Flash	STM32Cube.AI version
st_yolo_lc_v1	Int8	192x192x3	STM32H7	166.29	8.09	276.73	52.81	174.38	329.54	10.2.0
st_yolo_lc_v1	Int8	224x224x3	STM32H7	217.29	8.09	276.73	52.82	225.38	329.55	10.2.0
st_yolo_lc_v1	Int8	256x256x3	STM32H7	278.29	8.09	276.73	52.81	286.38	329.54	10.2.0

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

Model	Format	Resolution	Board	Execution Engine	Frequency	Inference time (ms)	STM32Cube.AI version
st_yolo_lc_v1	Int8	192x192x3	STM32H747I-DISCO	1 CPU	400 MHz	179.36	10.2.0
st_yolo_lc_v1	Int8	224x224x3	STM32H747I-DISCO	1 CPU	400 MHz	244.75	10.2.0
st_yolo_lc_v1	Int8	256x256x3	STM32H747I-DISCO	1 CPU	400 MHz	320.79	10.2.0

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

Model	Format	Resolution	Quantization	Board	Execution Engine	Frequency	Inference time (ms)	%NPU	%GPU	%CPU	X-LINUX-AI version	Framework
st_yolo_lc_v1	Int8	192x192x3	per-channel**	STM32MP257F-DK2	NPU/GPU	800 MHz	11.88 ms	2.62	97.38	0	v6.1.0	OpenVX
st_yolo_lc_v1	Int8	224x224x3	per-channel**	STM32MP257F-DK2	NPU/GPU	800 MHz	17.60 ms	3.33	96.67	0	v6.1.0	OpenVX
st_yolo_lc_v1	Int8	256x256x3	per-channel**	STM32MP257F-DK2	NPU/GPU	800 MHz	13.93 ms	5.12	94.88	0	v6.1.0	OpenVX
st_yolo_lc_v1	Int8	192x192x3	per-channel	STM32MP157F-DK2	2 CPU	800 MHz	33.38 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
st_yolo_lc_v1	Int8	224x224x3	per-channel	STM32MP157F-DK2	2 CPU	800 MHz	45.43 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
st_yolo_lc_v1	Int8	256x256x3	per-channel	STM32MP157F-DK2	2 CPU	800 MHz	58.80 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
st_yolo_lc_v1	Int8	192x192x3	per-channel	STM32MP135F-DK2	1 CPU	1000 MHz	52.63 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
st_yolo_lc_v1	Int8	224x224x3	per-channel	STM32MP135F-DK2	1 CPU	1000 MHz	72.51 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
st_yolo_lc_v1	Int8	256x256x3	per-channel	STM32MP135F-DK2	1 CPU	1000 MHz	95.84 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0

** To get the most out of MP25 NPU hardware acceleration, please use per-tensor quantization

AP on COCO Person dataset

Dataset details: link , License CC BY 4.0 , Quotation[1] , Number of classes: 80, Number of images: 118,287

Model	Format	Resolution	AP
st_yolo_lc_v1	Int8	192x192x3	30.7 %
st_yolo_lc_v1	Float	192x192x3	31.2 %
st_yolo_lc_v1	Int8	224x224x3	34.2 %
st_yolo_lc_v1	Float	224x224x3	33.8 %
st_yolo_lc_v1	Int8	256x256x3	35.6 %
st_yolo_lc_v1	Float	256x256x3	36.4 %

* 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

[1] “Microsoft COCO: Common Objects in Context”. [Online]. Available: https://cocodataset.org/#download. @article{DBLP:journals/corr/LinMBHPRDZ14, author = {Tsung{-}Yi Lin and Michael Maire and Serge J. Belongie and Lubomir D. Bourdev and Ross B. Girshick and James Hays and Pietro Perona and Deva Ramanan and Piotr Doll{'{a} }r and C. Lawrence Zitnick}, title = {Microsoft {COCO:} Common Objects in Context}, journal = {CoRR}, volume = {abs/1405.0312}, year = {2014}, url = {http://arxiv.org/abs/1405.0312}, archivePrefix = {arXiv}, eprint = {1405.0312}, timestamp = {Mon, 13 Aug 2018 16:48:13 +0200}, biburl = {https://dblp.org/rec/bib/journals/corr/LinMBHPRDZ14}, bibsource = {dblp computer science bibliography, https://dblp.org} }