Is it possible to use some of the openMV features such as output tensor processing (as mentioned with the latest v4.5.6 post with numpy processing the outputs) from inputs being sent to an external chip with NPU capabilities ? How would we integrate this ? Are there any examples, what would a typical flow be (external chip runs the network)
2nd part is unrelated to the first
If I purely want to use the SPI bus to transfer pre-processed input images to another chip (RISCV chip) which of the products would be best suited for this for use as a SPI Slave (using rpc)
Is it possible to use some of the openMV features such as output tensor processing (as mentioned with the latest v4.5.6 post with numpy processing the outputs) from inputs being sent to an external chip with NPU capabilities ? How would we integrate this ? Are there any examples, what would a typical flow be (external chip runs the network)
If you want to do something like this you would need to figure it out yourself. However, it’s no hard, you can directly turn ndarrays into byte arrays and then transmit those using any MicroPython function that supports the buffer protocol. Then, because the output is an ndarray you can manipulate the array in anyway you need in Python.
If I purely want to use the SPI bus to transfer pre-processed input images to another chip (RISCV chip) which of the products would be best suited for this for use as a SPI Slave (using rpc)
Use the OpenMV Cam H7 Plus for fast SPI slave/master support. The RT1062 doesn’t have the right software package via the machine module for this.
If you want to do something like this you would need to figure it out yourself. However, it’s no hard, you can directly turn ndarrays into byte arrays and then transmit those using any MicroPython function that supports the buffer protocol. Then, because the output is an ndarray you can manipulate the array in anyway you need in Python.
Ok openmv image is captured, pre-processed, converted to bytearray and transmitted. external chip with NPU runs the network, provides an an output ndarray. OpenMV processes the output tensor, perhaps shows a bounding box if its object detection, What about comms ? Do we use SPI/RPC ?