I am participating in the Sony Spresense Hackster.io competition and given board has 200 KB memory, i was looking to use Edge impulse trained model for my APP.
When I trained model, it was 30% accurate, so I ended up writing my own Yolo based Model with 95% accuracy.
Final step is to put this on the board, but given its size which is in GB now, I need to compress using Edge Impulse API, but I am not able to add it as custom block so I can transfer on to Sony Board
I only have to do this once, but there is no option to pay and get this kind of access as a developer?
How can I import my custom transfer learning blocks into Edge Impulse. It says that the feature is only able for Enterprise customers. Can I as a developer use the same feature and implement it into my project?
I believe you will never be able to run a Yolo-based model on the Spresense any time soon due to the resources constraints. The only object detection model that would fit on this board are FOMO models (I guess with small images, I would start with 64x64 images and if it works increase slightly the resolution if needed).
@louis thanks for confirmation. I trained my own YOLOv4 model to detect walrus and deployed it on Jetson Nano. It managed to get 95% accuracy.
Thanks for support.