Object Detection low accuracy

Thank you for awesome Edge Impulse Object Detection Block.
It is awesome block! I checked the model works in both of RPI 4 and mobile phone.

But I meet a problem. I want to ask two questions.

  1. The accuracy of the model is too low. I there any advice to upgrade the accuracy?
    My public project number is 26357

  2. I guess the cause of the accuracy drop is related to the resolution of the webcam connected to RPI4 and the resolution of the mobile phone.

The input data of the model was collected by a high-definition mobile phone. As shown in the table below, the resolution of the scrunch is different from that of the mobile phone and the webcam of RPI 4. When I tested the real model on RPI 4, it didn’t catch the scrunch well.


@dansitu Here performance on test set is indeed very low (also incredibly high drop in accuracy for the quantized model), any idea on how to improve? Longer training cycles? More varied data? I’ve added you to the project.