Is coral support likely?

Hi, new to deepstack and loving it so far, thanks for your work.

I’ve searched for mention of the google coral devices, and wonder if support for them is on your roadmap at all?

Thanks

1 Like

RPi version does not require hardware accel e.g. coral or NCS. If you really need high FPS checkout Frigate

1 Like

Hi @robmarkcole , any update on coral tpu integration , im using deepstack for face recognition and frigate+coral for person detection on my synology nas using docker , is there a way to use coral tpu to get lesser cpu loads and faster inference speed.

Thanks

So you want to offload face detection on TPU? Im quite surprised frigate doesn’t offer that yet

Frigate doesn’t have any detection currently, just basic support of coco object detection. I suspect more AI detection will come in the future, it doesn’t appear that’s the main focus of Frigate at the moment, which is mostly focused on making a good NVR (file storage and such)

There’s another outside service called Double Take https://github.com/jakowenko/double-take which integrates with Frigate, MQTT and Home-assistant, but currently it offloads the AI detection to other pipelines, which it recommends deepstack first. Because of this, you’re going to end up getting people from Frigate, which this community mostly uses Coral TPUs.

As for inference speed, this isn’t the only reason to offload to a TPU. There’s also inference per watt (low power usage) and freeing up CPU cycles on say a NAS or server to perform other tasks.

I personally use the TPU on a dell r720 which has 32 cores. While I could do inference on the CPUs, I’d live those CPU cycles free’d up to perform other tasks. Also if these CPUs are pinned performing AI tasks, the server uses 200watts of power, where if I could offload to TPU it would only add 1-2 watts of power. That reduces not only power, but heat and noise.

@j0rd checkout https://github.com/robmarkcole/coral-pi-rest-server

IMO deepstack will not be supporting coral hardware as this would be a complete nightmare for maintainers. For your use case I think adding features to frigate is the way to go, e.g. custom detectors

1 Like