STMicroelectronics/tiny_yolo_v2

Tiny Yolo v2 quantized

Use case : Object detection

Model description

Tiny Yolo v2 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
Provenance	https://github.com/AlexeyAB/darknet
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]	[]
STM32MP1	[x]	[x]
STM32MP2	[x]	[x]
STM32N6	[x]	[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)	External RAM (KiB)	Weights Flash (KiB)	STM32Cube.AI version	STEdgeAI Core version
tiny_yolo_v2	COCO-Person	Int8	224x224x3	STM32N6	392	0	10794.4	10.2.0	2.2.0
tiny_yolo_v2	ST-Person	Int8	224x224x3	STM32N6	392	0	10794.4	10.2.0	2.2.0
tiny_yolo_v2	COCO-Person	Int8	416x416x3	STM32N6	1880.12	0	10818.4	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
tiny_yolo_v2	COCO-Person	Int8	224x224x3	STM32N6570-DK	NPU/MCU	30.68	32.59	10.2.0	2.2.0
tiny_yolo_v2	ST-Person	Int8	224x224x3	STM32N6570-DK	NPU/MCU	30.68	32.59	10.2.0	2.2.0
tiny_yolo_v2	COCO-Person	Int8	416x416x3	STM32N6570-DK	NPU/MCU	50.81	19.68	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
tiny_yolo_v2	Int8	192x192x3	STM32H7	220.6 KiB	7.98 KiB	10775.98 KiB	55.85 KiB	228.58 KiB	10831.83 KiB	10.2.0
tiny_yolo_v2	Int8	224x224x3	STM32H7	249.35	7.98 KiB	10775.98	55.24	257.33	10831.22	10.2.0
tiny_yolo_v2	Int8	416x416x3	STM32H7	1263.07	8.03 KiB	10775.98	55.29	1271.10	10831.27	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
tiny_yolo_v2	Int8	192x192x3	STM32H747I-DISCO	1 CPU	400 MHz	3006.3 ms	10.2.0
tiny_yolo_v2	Int8	224x224x3	STM32H747I-DISCO	1 CPU	400 MHz	2712.22	10.2.0
tiny_yolo_v2	Int8	416x416x3	STM32H747I-DISCO	1 CPU	400 MHz	10441.33	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
tiny_yolo_v2	Int8	224x224x3	per-channel**	STM32MP257F-DK2	NPU/GPU	800 MHz	137.51 ms	3.83	96.17	0	v6.1.0	OpenVX
tiny_yolo_v2	Int8	416x416x3	per-channel**	STM32MP257F-DK2	NPU/GPU	800 MHz	481.09 ms	2.57	97.43	0	v6.1.0	OpenVX
tiny_yolo_v2	Int8	224x224x3	per-channel	STM32MP157F-DK2	2 CPU	800 MHz	417.84 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
tiny_yolo_v2	Int8	416x416x3	per-channel	STM32MP157F-DK2	2 CPU	800 MHz	1302.07 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
tiny_yolo_v2	Int8	224x224x3	per-channel	STM32MP135F-DK2	1 CPU	1000 MHz	623.35 ms	NA	NA	100	v6.1.0	TensorFlowLite 2.18.0
tiny_yolo_v2	Int8	416x416x3	per-channel	STM32MP135F-DK2	1 CPU	1000 MHz	2101.3 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
tiny_yolo_v2	Int8	192x192x3	26.2 %
tiny_yolo_v2	Float	192x192x3	27.5 %
tiny_yolo_v2	Int8	224x224x3	28.8 %
tiny_yolo_v2	Float	224x224x3	30.9 %
tiny_yolo_v2	Int8	416x416x3	41.4 %
tiny_yolo_v2	Float	416x416x3	43.4 %

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

AP on ST Person dataset

This model has been trained using a STMicroelectronics proprietary dataset which is not provided as part of the STM32 model zoo. The ST person dataset has been built by aggregating several public datasets and by applying data augmentation on these public datasets. If users wish to retrain this model it has to be done using another dataset selected by the user.

Model	Format	Resolution	AP
tiny_yolo_v2	Int8	224x224x3	28.5 %

* 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} }