Custom Training Help

[ahndroo]

I'm training on a 1 class dataset that has relatively small objects. When training I noticed that the loss on Output 3 is huge compared to the others but eventually seemingly converges. However, when running the trained model output, the bounding boxes are just wrong. Both the xmin,xmax and ymin,ymax coordinates are the same in some cases, or they define almost the entire image.

Has anyone encountered something like this?

[xiongliu666]

我也遇到了这个问题，我正在训练安全帽预测问题，训练后运行效果很差，边界框很多错误。