Mask Prediction Issue with Large Objects in Custom Dataset Training

[ctrimborn]

Hi,

I am training on a custom 1920 dataset. The objects are quite large (80%-90% of the width and height of the image). However, when I check if the model was trained successfully, the mask is always in the same spot (roughly the middle of the image) with a maximum width of 900 px.

Interestingly, when I lower the resolution to 920 or 640, the objects are recognized normally. The validation results show good performance, as seen in the image (Mask P: 0.984, R: 0.979, mAP50: 0.993, mAP50-95: 0.823), but the issue with object positioning at the higher resolution remains.

Could this be related to the mosaic augmentation making the objects appear smaller than they are in the original images? I'm training with the --rect parameter, so there’s no distortion in the pictures, but I can’t figure out why the model is showing this behavior.

If anyone can help, it would be much appreciated.

Kind regards,
Christian