I’m trying to wrap my head around a problem I’m experiencing. I know how to train image classification models and deploy them on a STM32H747-DISCO board, based on a food classification project. When I want to go for a object detection model I get stuck.
I’ve trained a model with Edge Impulse and I get a giant CMSIS pack… my question is, where do I go from there? Does anyone have any experience with object detection models on stm32h7 boards?
ST does have a people counting project based on a object detection model, but I can’t modify that model.
I have exactly the same problem, so I’m going to revive the post.
I would like to perform object detection (with bounding box) on an STM32H747I-Disco, once the model has been trained and the .pack file downloaded, is there an example Cube IDE project in which computer vision is performed with the Edge Impulse libraries or do you know how to modify the PeopleCounting project of the FP-AI-VISION1 pack of ST in order to replace their compiled libraries with the one given by Edge Impulse?
Constrained object detection has been released since the start of this thread. I recommend taking a look at this presentation (tinyML Talks: Constrained Object Detection on Microcontrollers with FOMO - YouTube) to learn more about how FOMO works and its limitations. It won’t give you a true bounding box. Rather, it tries to figure out which “grid cells” in your image contains your object(s) in question.
We don’t have any examples for the STM32H7, just the documentation showing how to use the .pack file here: Cube.MX CMSIS-PACK - Edge Impulse Documentation. In theory, it should work the same (just different pins). If you get it working with the H7, please let us know!
