I could wrote with parts from here and there my own code, without know absoluty nothing of coding.
What I got:
YOLO 20 CLASSES maix example but with all the features of StickV, as turn off,torch,etc. Draw the boundary boxes with better size and more clear bounds. Send via USB the coordinates,fps,class of the bboxes. Maix example only give FPS
Any MOBILENET example code as 75 classes, or my own trained model keeping the features of the original firmware as Torch,buttons etc, as if you train your own model you lose that things. Send via USB Class of the object, fps and size of the inference (I think but not sure, I print kpu.foward (task, img) )
My trained model recognize all the kind of surface of my house, parquet, carpet,tiles, etc my rabbit and people. It works very nice. It is to pass that info to my robots.
The think is I dont know what protocol use via USB, I think is Uart via VCP, but not sure. I would like to know the grove connector if is programble and how, pins 34 35, it should be possible, but no idea for what is used, if is fixed to I2C, SPI or what.
Now what I want to know too is how get bbox in mobilenet models, it only give text, but dont draw boxes.
Example of UART there is in the chinese version of MAIXPY wiki/ API.
The web of training is just amazing and worth more that any support from MAIX, M5Stack should not leave the support in hands of Sipeed, their support and realibility is null.
https://www.instructables.com/id/Object-Detection-With-Sipeed-MaiX-BoardsKendryte-K/
Follow this guy, he helped more that the own company, and check out all possible OPENMV IDE and forums, that is from MAIXPY was forked