object_detection/docs/README_MODELS_TORCH.md

Overview of PyTorch STM32 Model Zoo for Object Detection

The STM32 model zoo includes several PyTorch-based models for object detection use cases, converted to ONNX format and optimized for STM32N6 NPU deployment. All models are pre-trained on ImageNet and quantized using QDQ INT8 quantization.

Model Categories

• ST_pretrainedmodel_public_dataset contains PyTorch object detection models pretrained by ST on open source datasets such as MS COCO, Pascal VOC, COCO person and converted to ONNX format with QDQ quantization.

Explore the complete PyTorch model zoo with pre-trained models optimized for STM32N6.

Model Families

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

Single Shot Detector (SSD) Models

Variation of SSD models using different backbone and heads

• ssd_mobilenetv1_pt – Single shot detector with MobilenetV1 backbone
• ssd_mobilenetv2_pt – Single shot detector with MobilenetV2 backbone
• ssdlite_mobilenetv1_pt – Single shot detector lite with MobilenetV1 backbone
• ssdlite_mobilenetv2_pt – Single shot detector lite with MobilenetV2 backbone
• ssdlite_mobilenetv3mall_pt – Single shot detector lite with MobilenetV3small backbone
• ssdlite_mobilenetv3large_pt – Single shot detector lite with MobilenetV3large backbone

ST_YOLOD Models

STMicroelectronics in house developed model especially optimized for size and memory with a competitive performance on Imagenet (and other datasets)

• st_yolodv2milli_pt – Extreme compression of STResNet for backbone and YOLOX head and modified YOLOX neck.
• st_yolodv2tiny_pt – Extreme compression of STResNet for backbone and YOLOX head and modified YOLOX neck.

Feel free to explore the model zoo and get pre-trained models here. For training and deployment guidance, refer to the STM32 AI model zoo documentation.