3.1 — ST Model Zoo

3.1 — ST Model Zoo

Where the models come from

The ST Model Zoo is a public GitHub repository maintained by STMicroelectronics containing a curated collection of pre-trained neural network models, all validated for deployment on STM32 devices.

🔗 github.com/STMicroelectronics/stm32ai-modelzoo

What it is — and why it matters

Training a neural network from scratch requires large datasets, powerful GPUs, and days of computation. For most embedded AI projects — including this one — this is neither necessary nor practical. The Model Zoo solves this by providing models that are already trained, quantized, and validated on reference STM32 hardware. You download the model file and proceed directly to deployment.

As of release 4.0, the Model Zoo supports three frameworks — TensorFlow, PyTorch, and ONNX — and provides performance benchmarks on both MCU and MPU targets for every model.

What’s new in release 4.0 (latest):

• Major PyTorch support for Image Classification and Object Detection
• Support of STEdgeAI Core v3.0.0
• New use cases: Face Detection, Arc Fault Detection, Re-Identification
• Mixed precision models (Weights 4-bit, Activations 8-bit)
• Support for Keras 3.8.0, TensorFlow 2.18.0, PyTorch 2.7.1, ONNX 1.16.1
• Docker-based setup with ready-to-use image including the full software stack

Repository structure

The repository is organised by application domain. Each folder contains multiple model architectures, each pre-trained on public datasets. Within each domain, models are further organised by training source.

stm32ai-modelzoo/
├── arc_fault_detection/
├── audio_event_detection/
├── depth_estimation/
├── face_detection/
├── hand_posture/
├── human_activity_recognition/
├── image_classification/
├── instance_segmentation/
├── neural_style_transfer/
├── object_detection/
├── pose_estimation/    ← used in this project
│   ├── handlandmarks/
│   ├── headlandmarks/
│   ├── movenet/        ← MoveNet Lightning lives here
│   │   ├── ST_pretrainedmodel_public_dataset/
│   │   ├── ST_pretrainedmodel_custom_dataset/
│   │   └── Public_pretrainedmodel_custom_dataset/
│   ├── yolov8n_pose/
│   └── yolov11n_pose/
├── re_identification/
├── semantic_segmentation/
└── speech_enhancement/

Subfolder	Description
ST_pretrainedmodel_public_dataset	Models trained by ST on public datasets (e.g. COCO). Best validated option.
ST_pretrainedmodel_custom_dataset	Models trained by ST on custom / proprietary datasets.
Public_pretrainedmodel_public_dataset	Public models (e.g. Google MoveNet) on public datasets.

Available use cases

The table below shows all supported use cases and which STM32 board is recommended for each. Notice that the STM32N6570-DK supports the widest range of use cases — a direct consequence of its Neural-ART NPU.

Use Case	Description	STM32N6570-DK
Image Classification	Classifies content within predefined classes	✓
Object Detection	Detects and locates objects in images	✓
Pose Estimation ← this project	Detects keypoints on people (hands, face, body)	Recommended
Semantic Segmentation	Labels every pixel in an image	✓
Instance Segmentation	Labels pixels per object instance	✓
Depth Estimation	Pixel-wise depth map from image	✓
Neural Style Transfer	Applies artistic style to images	✓
Speech Enhancement	Audio enhancement in noisy environments	✓
Re-Identification	Reidentifies person/object across frames	✓

Zoom in: Pose Estimation folder

The pose_estimation/ folder contains five model families:

• movenet — single pose, 13 or 17 keypoints  used in this project
• yolov8n_pose — multi pose, 17 keypoints (COCO standard)  used in this project
• yolov11n_pose — multi pose, 17 keypoints (latest YOLO)
• handlandmarks — single pose, 21 hand keypoints
• headlandmarks — single pose, 468/478 face keypoints

MoveNet Lightning — Official performance on STM32N6570-DK

The following data comes directly from the Model Zoo README — these are the official ST benchmarks for the model we deployed. Our measured results (22 ms at 192×192) are consistent with these figures.

Memory footprint on STM32N6

Model	Resolution	Format	Internal RAM (KiB)	Weights Flash (KiB)
ST MoveNet Lightning heatmaps	192×192×3	Int8	914.88	2304.0
ST MoveNet Lightning heatmaps	224×224×3	Int8	1239.04	2304.0
ST MoveNet Lightning heatmaps	256×256×3	Int8	1607.68	2304.0

Inference time on STM32N6570-DK (NPU/MCU)

Model	Resolution	Inference (ms)	Inf/sec	STEdgeAI version
ST MoveNet Lightning heatmaps	192×192×3	22.05 ms	45.35	3.0.0
ST MoveNet Lightning heatmaps	224×224×3	27.64 ms	36.18	3.0.0
ST MoveNet Lightning heatmaps	256×256×3	35.50 ms	28.17	3.0.0

Accuracy — OKS on COCO Person dataset

Model	Resolution	Quantization	OKS (%)
ST MoveNet Lightning heatmaps	192×192	per-channel	57.64%
ST MoveNet Lightning heatmaps	224×224	per-channel	62.29%
ST MoveNet Lightning heatmaps	192×192	per-tensor	55.84%
MoveNet Lightning (Google)	192×192	per-channel	54.12%

How to clone and use it

The repo uses Git LFS for large binary files (model weights). You must install and initialise Git LFS before cloning, otherwise model files will download as empty pointers.

# 1. Install Git LFS (Ubuntu)
sudo apt-get install git-lfs

# 2. Initialise Git LFS for your account
git lfs install
# Expected output: Git LFS initialized.

# 3. Clone — place it in the SAME folder as modelzoo-services
git clone https://github.com/STMicroelectronics/stm32ai-modelzoo.git

# 4. Navigate to the model you want
cd stm32ai-modelzoo/pose_estimation/movenet/ST_pretrainedmodel_public_dataset/

Important — folder placement: Clone the Model Zoo in the same parent folder as stm32ai-modelzoo-services. The config file examples in Services reference model paths relative to this structure. If the folders are not siblings, you will need to update every model_path in user_config.yaml manually.

What we used in this project

Model	Source	Format	Zoo?	Quantization needed?
MoveNet Lightning	ST Model Zoo	.tflite INT8	Yes	No — pre-quantized
YOLOv8n-pose	Ultralytics (external)	.pt → .tflite	No	Yes — PTQ via TFLite
TinyBERT	HuggingFace (external)	.onnx	No	Yes — PTQ required

Next step: once you have your model file, you pass it to ModelZoo Services via user_config.yaml. The Services framework handles everything else — quantization if needed, profiling, and deployment. That is what we look at in the next section.

← Chapter 3 Overview Next: 3.2 ModelZoo Services →