YOLO support

OpenMV Products
1

Hi there,

I’m working in a project on which I need to detect people, and know where they are (in the image frame). I was using the on-board person detection model with many different ROIs (to determine where in the image the person is located). I tried using the Haar-Cascade face detector, but this performed worst than the multiple ROI approach.

Although the multiple-ROI algorithm works, this approach has some drawbacks, for example, if a person is between 2 adjacent ROIs, both of the ROIs appear to have a person inside. There is probably some workaround this, but I think a detector model like YOLO would best.

Is YOLO support comming anytime soon? Is ther a way I can load a YOLO model into a H7 plus cam?

I was previously using a K210 board which does support YOLO (Object Detection with Sipeed MaiX Boards (Kendryte K210) - Hackster.io), but I changed to the OpenMV cam since is so much better documented, and easy to use for prototypes.

Thanks a lot!

2

We have asked Edge Impulse to develop this. It’s very requested by folks.

To note, we’re not going to just support YOLO… this would not be useful. Instead, we want a whole training pipeline so you can train a CNN to do what you actually need.

I’ll send your request to Edge Impulse. Note that we want to support Segmentation Networks and Object Detection networks. Once they have support for this it will definitely be a game changer.