Overview of pose estimation STM32 model zoo

January 22, 2026 · View on GitHub

The STM32 model zoo includes several models for object detection use case pre-trained on custom and public datasets. Under each model directory, you can find the following model categories:

Public_pretrainedmodel_custom_dataset contains public pose estimation models trained on custom datasets.
ST_pretrainedmodel_custom_dataset contains different pose estimation models trained on ST custom datasets.
ST_pretrainedmodel_public_dataset contains pose estimation models trained on public datasets.

Feel free to explore the model zoo and get pre-trained models here.

You can get footprints and performance information for each model following links below:

To get started, update the user_config.yaml file, which specifies the parameters and configuration options for the services you want to use. The model section of this yaml specifically relates to the model definition. The model_type is mandatory and must correspond to your model. Some topologies are already registered and can be accessed via the model_name attribute.

`model_name`

The exhaustive list of possible model_name is provided hereafter:

'custom_models'
'st_movenet_lightning_a100_heatmaps'
'st_movenet_lightning_heatmaps'
- alpha = value between 0.35 & 1.4, (but only [0.35, 0.5, 0.75, 1.0 ,1.4] come with pretrained ImageNet weights)

Important

'st_movenet_lightning_heatmaps' model has a mandatory alpha parameter that controls the width of the network. This is known as the width multiplier in the MobileNetV2 paper.

alpha < 1.0 : proportionally decreases the number of filters in each layer.
alpha > 1.0 : proportionally increases the number of filters in each layer.
alpha == 1 : default number of filters from the paper are used at each layer -> equivalent to 'st_movenet_lightning_a100_heatmaps'

`model_type`

The model_type attribute specifies the type of the model architecture that you want to use, it represents a specific post-processing / output shape / use-case.

heatmaps_spe: These are single pose estimation models that outputs heatmaps that we must post-process in order to get the keypoints positions and confidences.
spe: These are single pose estimation models that output directly the keypoints positions and confidences.
yolo_mpe : These are the YOLO (You Only Look Once) multiple pose estimation models from Ultralytics that outputs the same tensor as in object detection but with the addition of a set of keypoints for each bbox.
hand_spe: These are single hand landmarks estimation models that outputs directly the keypoints positions and confidences of the hand pose.
head_spe: These are single head landmarks estimation models that outputs directly the keypoints positions and confidences of the head pose.

Summary

Models with their corresponding model_name, model_type, keypoints, class_names & supported services:

model	`model_name`	`model_type`	`keypoints`	`class_names`	Training	Evaluation	Prediction	Deployment
st_movenet_lightning	`st_movenet_lightning_a100_heatmaps`, `st_movenet_lightning_heatmaps`	`heatmaps_spe`	`17`	`[person]`	✅	✅	✅	✅
Custom model	`custom_models`	`heatmaps_spe`	customizable	customizable	✅	✅	✅	✅
movenet_lightning	Not registered*	`spe`	`17`	`[person]`	❌	✅	✅	✅
hand_landmarks	Not registered*	`hand_spe`	`21`	`[hand]`	❌	❌	✅	✅
head_landmarks	Not registered*	`head_spe`	`468`	`[head]`	❌	❌	✅	✅
yolo_v8n	Not registered*	`yolo_mpe`	`17`	`[person]`	❌	✅	✅	✅
yolo_v11n	Not registered*	`yolo_mpe`	`17`	`[person]`	❌	✅	✅	✅

* Use the model_path instead