STYOLOTiny
Use case : Object detection
Model description
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).
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.
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.
Network information
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, (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 |
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 |
|---|
| st_yolodv2tiny_pt | COCO | Int8 | 192x192x3 | STM32N6 | 1597.50 | 0 | 4789.56 | 3.0.0 |
| st_yolodv2tiny_pt | COCO | Int8 | 288x288x3 | STM32N6 | 2644.00 | 648.00 | 4801.37 | 3.0.0 |
| st_yolodv2tiny_pt | COCO | Int8 | 640x640x3 | STM32N6 | 2796.56 | 9600.00 | 4811.48 | 3.0.0 |
- 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.
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 |
|---|
| st_yolodv2tiny_pt | COCO | Int8 | 192x192x3 | STM32N6570-DK | NPU/MCU | 26.52 | 37.71 | 3.0.0 |
| st_yolodv2tiny_pt | COCO | Int8 | 288x288x3 | STM32N6570-DK | NPU/MCU | 66.66 | 15.00 | 3.0.0 |
| st_yolodv2tiny_pt | COCO | Int8 | 640x640x3 | STM32N6570-DK | NPU/MCU | 3100.00 | 0.32 | 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 |
|---|
| st_yolodv2tiny_pt | COCO-Person | Int8 | 192x192x3 | STM32N6 | 1597.50 | 0 | 4767.34 | 3.0.0 |
| st_yolodv2tiny_pt | COCO-Person | Int8 | 256x256x3 | STM32N6 | 2424.00 | 0 | 4774.09 | 3.0.0 |
| st_yolodv2tiny_pt | COCO-Person | Int8 | 288x288x3 | STM32N6 | 2644.00 | 648.00 | 4779.15 | 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 |
|---|
| st_yolodv2tiny_pt | COCO-Person | Int8 | 192x192x3 | STM32N6570-DK | NPU/MCU | 25.50 | 39.22 | 3.0.0 |
| st_yolodv2tiny_pt | COCO-Person | Int8 | 256x256x3 | STM32N6570-DK | NPU/MCU | 37.23 | 26.86 | 3.0.0 |
| st_yolodv2tiny_pt | COCO-Person | Int8 | 288x288x3 | STM32N6570-DK | NPU/MCU | 64.87 | 15.42 | 3.0.0 |
AP on COCO dataset
Dataset details:
link , License
CC BY 4.0, Number of classes: 80
* 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
* 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
hereReferences
