Hugging Face's logo

STMicroelectronics
/
st_yolodv2tiny_pt

Object Detection
Model card Files
xet
Community
FBAGSTM commited on
Commit
85d611d
·
verified ·
1 Parent(s): c6ad005

Release AI-ModelZoo-4.0.0

Browse files
Files changed (1) hide show
  1. README.md +144 -6
README.md CHANGED
@@ -1,6 +1,144 @@
1
- ---
2
- license: other
3
- license_name: sla0044
4
- license_link: >-
5
- https://github.com/STMicroelectronics/stm32ai-modelzoo/blob/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/LICENSE.md
6
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: sla0044
4
+ license_link: >-
5
+ https://github.com/STMicroelectronics/stm32ai-modelzoo/blob/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/LICENSE.md
6
+ pipeline_tag: object-detection
7
+ ---
8
+ # **STYOLOTiny**
9
+
10
+ ## **Use case** : `Object detection`
11
+
12
+ ## **Model description**
13
+
14
+ STYOLOTiny is a **compact, efficient object detection model** designed for deployment on **resource‑constrained devices** such as microcontrollers (MCUs), embedded systems, and edge accelerators. It belongs to the STYOLO family introduced in *“STResNet & STYOLO: A New Family of Compact Classification and Object Detection Models for MCUs”* (arXiv:2601.05364).
15
+
16
+ The model uses a **highly compressed backbone** derived from STResNet along with a **YOLO‑style detection head**, optimized for **low memory usage, minimal compute, and real‑time inference**. STYOLOTiny targets scenarios where **small model footprint and fast inference** are critical, offering competitive detection performance on standard benchmarks compared to other micro‑sized detectors.
17
+
18
+ The `st_yolodv2tiny_pt` variant has a slightly larger footprint compared to `st_yolodv2milli_pt` and it is implemented in **PyTorch** and intended for **ultra‑efficient object detection** on edge and MCU environments.
19
+
20
+
21
+ ## **Network information**
22
+
23
+ | Network information | Value |
24
+ |--------------------|-------|
25
+ | Framework | Torch |
26
+ | Quantization | Int8 |
27
+ | Provenance | [STMicroelectronics Model Zoo Services](https://github.com/STMicroelectronics/stm32ai-modelzoo-services/tree/main/object_detection/object_detection) |
28
+ | Paper | [STResNet & STYOLO (arXiv:2601.05364)](https://arxiv.org/abs/2601.05364) |
29
+
30
+ The model is quantized to **int8** using **ONNX Runtime** and exported for efficient deployment.
31
+
32
+
33
+ ## Network inputs / outputs
34
+
35
+ For an image resolution of NxM and NC classes
36
+
37
+ | Input Shape | Description |
38
+ | ----- | ----------- |
39
+ | (1, W, H, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
40
+
41
+ | Output Shape | Description |
42
+ | ----- | ----------- |
43
+ | (1, (W/8xH/8 + W/16xH/16 + W/32xH/32), (NC+1+4)) | Model returns bounding boxes with 6 values for each box, four coordinates (x1,y1,x2,y2), class confidence and objectness confidence |
44
+
45
+
46
+ ## Recommended Platforms
47
+
48
+ | Platform | Supported | Recommended |
49
+ |----------|-----------|-------------|
50
+ | STM32L0 | [] | [] |
51
+ | STM32L4 | [] | [] |
52
+ | STM32U5 | [] | [] |
53
+ | STM32H7 | [] | [] |
54
+ | STM32MP1 | [] | [] |
55
+ | STM32MP2 | [] | [] |
56
+ | STM32N6 | [x] | [x] |
57
+
58
+
59
+ # Performances
60
+
61
+ ## Metrics
62
+
63
+ Measures are done with default STEdgeAI Core configuration with enabled input / output allocated option.
64
+
65
+ ### Reference **NPU** memory footprint based on COCO dataset (see Accuracy for details on dataset)
66
+ | Model | Dataset | Format | Resolution | Series | Internal RAM (KiB) | External RAM (KiB) | Weights Flash (KiB) | STEdgeAI Core version |
67
+ |-------|---------|--------|------------|--------|-------------------|-------------------|--------------------|-----------------------|
68
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_192/st_yolodv2tiny_actrelu_pt_coco_192_qdq_int8.onnx) | COCO | Int8 | 192x192x3 | STM32N6 | 1597.50 | 0 | 4789.56 | 3.0.0 |
69
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_288/st_yolodv2tiny_actrelu_pt_coco_288_qdq_int8.onnx) | COCO | Int8 | 288x288x3 | STM32N6 | 2644.00 | 648.00 | 4801.37 | 3.0.0 |
70
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_640/st_yolodv2tiny_actrelu_pt_coco_640_qdq_int8.onnx) | COCO | Int8 | 640x640x3 | STM32N6 | 2796.56 | 9600.00 | 4811.48 | 3.0.0 |
71
+
72
+ * 640x640 coco checkpoints are provided primarily for finetuning purposes as these checkpoints are trained on large dataset at higher resolution. Models with 640 resolution is not suitable for deployment.
73
+
74
+ ### Reference **NPU** inference time based on COCO dataset (see Accuracy for details on dataset)
75
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
76
+ |-------|---------|--------|------------|-------|------------------|--------------------|-----------|-----------------------|
77
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_192/st_yolodv2tiny_actrelu_pt_coco_192_qdq_int8.onnx) | COCO | Int8 | 192x192x3 | STM32N6570-DK | NPU/MCU | 26.52 | 37.71 | 3.0.0 |
78
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_288/st_yolodv2tiny_actrelu_pt_coco_288_qdq_int8.onnx) | COCO | Int8 | 288x288x3 | STM32N6570-DK | NPU/MCU | 66.66 | 15.00 | 3.0.0 |
79
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_640/st_yolodv2tiny_actrelu_pt_coco_640_qdq_int8.onnx) | COCO | Int8 | 640x640x3 | STM32N6570-DK | NPU/MCU | 3100.00 | 0.32 | 3.0.0 |
80
+
81
+ ### Reference **NPU** memory footprint based on COCO Person dataset (see Accuracy for details on dataset)
82
+ | Model | Dataset | Format | Resolution | Series | Internal RAM (KiB) | External RAM (KiB) | Weights Flash (KiB) | STEdgeAI Core version |
83
+ |-------|---------|--------|------------|--------|-------------------|-------------------|--------------------|-----------------------|
84
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_192/st_yolodv2tiny_actrelu_pt_coco_person_192_qdq_int8.onnx) | COCO-Person | Int8 | 192x192x3 | STM32N6 | 1597.50 | 0 | 4767.34 | 3.0.0 |
85
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_256/st_yolodv2tiny_actrelu_pt_coco_person_256_qdq_int8.onnx) | COCO-Person | Int8 | 256x256x3 | STM32N6 | 2424.00 | 0 | 4774.09 | 3.0.0 |
86
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_288/st_yolodv2tiny_actrelu_pt_coco_person_288_qdq_int8.onnx) | COCO-Person | Int8 | 288x288x3 | STM32N6 | 2644.00 | 648.00 | 4779.15 | 3.0.0 |
87
+
88
+ ### Reference **NPU** inference time based on COCO Person dataset (see Accuracy for details on dataset)
89
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
90
+ |-------|---------|--------|------------|-------|------------------|--------------------|-----------|-----------------------|
91
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_192/st_yolodv2tiny_actrelu_pt_coco_person_192_qdq_int8.onnx) | COCO-Person | Int8 | 192x192x3 | STM32N6570-DK | NPU/MCU | 25.50 | 39.22 | 3.0.0 |
92
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_256/st_yolodv2tiny_actrelu_pt_coco_person_256_qdq_int8.onnx) | COCO-Person | Int8 | 256x256x3 | STM32N6570-DK | NPU/MCU | 37.23 | 26.86 | 3.0.0 |
93
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_288/st_yolodv2tiny_actrelu_pt_coco_person_288_qdq_int8.onnx) | COCO-Person | Int8 | 288x288x3 | STM32N6570-DK | NPU/MCU | 64.87 | 15.42 | 3.0.0 |
94
+
95
+
96
+
97
+
98
+ ### AP on COCO dataset
99
+
100
+ Dataset details: [link](https://cocodataset.org/#download) , License [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/legalcode), Number of classes: 80
101
+
102
+ | Model | Format | Resolution | AP50 |
103
+ | --- | --- | --- | --- |
104
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_192/st_yolodv2tiny_actrelu_pt_coco_192.onnx) | Float | 3x192x192 | 37.45 |
105
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_192/st_yolodv2tiny_actrelu_pt_coco_192_qdq_int8.onnx) | Int8 | 3x192x192 | 35.78 |
106
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_288/st_yolodv2tiny_actrelu_pt_coco_288.onnx) | Float | 3x288x288 | 45.62 |
107
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_288/st_yolodv2tiny_actrelu_pt_coco_288_qdq_int8.onnx) | Int8 | 3x288x288 | 44.57 |
108
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_640/st_yolodv2tiny_actrelu_pt_coco_640.onnx) | Float | 3x640x640 | 54.32 |
109
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco/st_yolodv2tiny_actrelu_pt_coco_640/st_yolodv2tiny_actrelu_pt_coco_640_qdq_int8.onnx) | Int8 | 3x640x640 | 53.46 |
110
+
111
+ \* EVAL_IOU = 0.5, NMS_THRESH = 0.5, SCORE_THRESH = 0.001, MAX_DETECTIONS = 100
112
+
113
+ ### AP on COCO-Person dataset
114
+
115
+ Dataset details: [link](https://cocodataset.org/#download) , License [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/legalcode) , Number of classes: 1
116
+
117
+
118
+ | Model | Format | Resolution | AP50 |
119
+ | --- | --- | --- | --- |
120
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_192/st_yolodv2tiny_actrelu_pt_coco_person_192.onnx) | Float | 3x192x192 | 62.75 |
121
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_192/st_yolodv2tiny_actrelu_pt_coco_person_192_qdq_int8.onnx) | Int8 | 3x192x192 | 62.15 |
122
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_256/st_yolodv2tiny_actrelu_pt_coco_person_256.onnx) | Float | 3x256x256 | 69.22 |
123
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_256/st_yolodv2tiny_actrelu_pt_coco_person_256_qdq_int8.onnx) | Int8 | 3x256x256 | 68.41 |
124
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_288/st_yolodv2tiny_actrelu_pt_coco_person_288.onnx) | Float | 3x288x288 | 71.02 |
125
+ | [st_yolodv2tiny_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/object_detection/st_yolodv2tiny_pt/ST_pretrainedmodel_public_dataset/coco_person/st_yolodv2tiny_actrelu_pt_coco_person_288/st_yolodv2tiny_actrelu_pt_coco_person_288_qdq_int8.onnx) | Int8 | 3x288x288 | 70.72 |
126
+
127
+ \* EVAL_IOU = 0.5, NMS_THRESH = 0.5, SCORE_THRESH = 0.001, MAX_DETECTIONS = 100
128
+
129
+ ## Retraining and Integration in a simple example:
130
+
131
+ Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
132
+
133
+
134
+ ## References
135
+
136
+
137
+ - **STYOLO / STResNet paper**
138
+ [S. Sah & R. Kumar, *STResNet & STYOLO: A New Family of Compact Classification and Object Detection Models for MCUs*](https://arxiv.org/abs/2601.05364)
139
+
140
+ - **YOLOX (inspires STYOLO architecture)**
141
+ [Ge et al., *YOLOX: Exceeding YOLO Series in 2021*](https://arxiv.org/abs/2107.08430)
142
+
143
+ - **MS COCO dataset**
144
+ [Lin et al., *Microsoft COCO: Common Objects in Context*](https://arxiv.org/abs/1405.0312)