Box detected take all the picture as one class....

[frobin02]

Hi, I trained my model class per class (one training image = one class / objet are 90 pourcent centred in each picture 640-640) and when i go to detect.py with my best.pt , the boxe select all my testing picture, with an approximate prediction when this same picture is composed by several classes to be detected ... , any ideas ?... it's possibly due to the training approche ? thanks

(PS; when testing picture is only zoomed on one classe, the model works very well but boxe take also all the picture...)