How to make XML for negative sample (background image) due to large error detection? I made XML for the negative samples and removed the object nodes. However, the trained model has no effect on the negative samples, and there is still a big error detection