We currently use ONNX Runtime for our object detection. We're working on replacing this (#105). However, it may take another week or two to optimize it to work good enough on Raspberry Pi.
So, for now, we should add it as feature-gated it to phone-as-a-camera (given the phone has more compute and we need a provider from Rust for AI ASAP), and, later on, when burn-flex is sufficiently optimized for Raspberry Pi as well, swap in Raspberry Pi.
Therefore, both should be supported in motion-ai, with phone-as-a-camera using burn-flex, and Raspberry Pi using ORT. Something like: "ort" or "flex" feature in the motion-ai library.
We currently use ONNX Runtime for our object detection. We're working on replacing this (#105). However, it may take another week or two to optimize it to work good enough on Raspberry Pi.
So, for now, we should add it as feature-gated it to phone-as-a-camera (given the phone has more compute and we need a provider from Rust for AI ASAP), and, later on, when burn-flex is sufficiently optimized for Raspberry Pi as well, swap in Raspberry Pi.
Therefore, both should be supported in motion-ai, with phone-as-a-camera using burn-flex, and Raspberry Pi using ORT. Something like: "ort" or "flex" feature in the motion-ai library.