Hi, i am working with OPENMV CAM H7+ for machine learning for image classification for fruit detection (16 lables). I created a CNN model with google Colab and it has great performance.
When i load the model (9 MB) on the board it works at 0.8 FPS and doesen’t recognise fruits properly.
Hi, the FPS of 0.8 is expected for a CNN of that size. Comparable models for image detection are in the 300KB range.
As for not working well, you need to have a huge amount of training data of what the camera has actually seen. If you trained the model on images from the web and then run it on camera images the model will suffer as it expects image that look like what is was trained on.
Please use Edge Impulse to train CNNs for embedded systems.
I understand the problem of FPS thank you. The model was trained on a dataset of 12,000 images divided into 16 labels, aren’t they enough?
Take some images from the camera and then classify them using your model on colab. Please let me know if there’s a bug difference is ability.