Using AMB82-Mini with a self-made model (yolov7 Pytorch) for object recognition performance is very poor

I completed the training using a custom model and uploaded the model weight file to
https://www.amebaiot.com/en/amebapro2-ai-convert-model/
After model conversion, the recognition performance burned into amb82 mini is extremely poor.

I tested darknet yolov3-tiny and yolov7 Pytorch respectively. When I confirmed the performance during training, the recognition performance was very good, and both mAP and accuracy were very high.

However, when I converted the training weights into models respectively, the recognition effect was very poor, and even classes other than my custom ones appeared (example: I only had two classes in training, so logically there should only be person and bicycle, but other classes appeared when I was identifying it)

I would like to know if anyone has the same problem as me and how to solve it.

1 Like