Replies: 4 comments 4 replies
-
@Curid I'm working on offering official benchmarks across most of our export pipelines here, but it's an enormous task. On the Edge TPU side, how do you benchmark? Which hardware exactly would be the 'reference' hardware, since Coral offers so many things? Also since these are merely accelerators I'm assuming the base system (CPU, RAM, storage) significantly influence results no? Do we need a Linux system for the USB accelerator or can we plug it into a Mac also? |
Beta Was this translation helpful? Give feedback.
-
To add some extra context, I want to produce official benchmarks for all our officially supported formats across any backend applicable, i.e. RPi4, Intel CPU, AMD CPU, M1 CPU, A15 ASIC (iOS), Snapdragon 888 (Android), V100 (GPU), Jetson (GPU), Coral (USB accelerator??) FormatsYOLOv5 inference is officially supported in 11 formats:
|
Beta Was this translation helpful? Give feedback.
-
I don't know what script they're using for benchmarking. I got the results from here.
They all use the same chip from what I understand.
I don't think so, the usb accelerator can be used with a rpi zero.
https://dev.to/kojikanao/use-coral-edgetpu-usb-accelerator-on-macos-435 |
Beta Was this translation helpful? Give feedback.
-
How does YOLOv5 compare to the Tensorflow models? I'm currently using SSDLite_MobileDet at 32.9 mAP and 9ms latency.
Does anyone know of any other Coral compatible person detection models besides my TF-CCTV?
Beta Was this translation helpful? Give feedback.
All reactions