GoAI 2.0 FAQ

The person detection model (Mobilenet v1) uses a 96x96 video input. Most models only need a small input image to work well. So, the 640x480 image needs to be downscaled to 96x96. This isn’t hard to update and there is already the code for doing this in the design (just not shown in the diagram). For the OV2640 camera we use, we place the camera in QQVGA mode which is 160x120 and then downscale to 96x96. Downscaling is just some counters to skip some pixels and lines. The model also works on monochrome images, so you’ll need to just keep the Y component of the incoming YCbCr data.
PSRAM and HyperRAM are essentially the same thing. We support HyperRAM and PSRAM for GoAI 2.0 layer memory and have the accelerator working on the GW1NSR4P.
As of now it’s about the same for larger devices. We’re still working on getting GW1NSR4 to the same speed or faster because we had to remove the flash DMA memory to get the design to fit. However, we have several updates that will improve the RAM throughput we are actively working on to get into the GW1NSR4.
It currently consumes almost all of the FPGA resources, but we are still doing optimization.
If there is an option to connect to the FPGA I would recommend it as we may switch to quad spi mode for improving the bandwidth soon.
No, Machine learning models can be developed to detect multiple outputs from the same model. This means that one GoAI machine learning processor can be used for multiple outputs.
No, Machine learning models are capable of inference on objects in images that the human eye cannot necessarily see. A larger model input does not necessarily improve the detection/inference result, yet does require more processing. For this reason, the input of image detection models are often times small.