关于 mosaic的使用prob

[hdmjdp]

DCIC22-SAR-ship-detection/configs/ship/cascade_swin_small_large_scale_onlytraindata_noflip_anchor4_ratio_mosaic.py

Line 212 in c44c29c

dict(type='Mosaic', img_scale=(256, 256), pad_val=0.0, prob=0.5),

大佬，你好，官方默认参数好像都是用1.0，你这里为啥是0.5。是不是对比过，0.5更好啊？

[zhengye1995]

Yes, 0.5 is better than 1.0 in this task.

[hdmjdp]

why not use mixup？

[zhengye1995]

why not use mixup？

I have tried it but the performance drops a lot, I gusses mixup two SAR images is not in line with its physical imaging principles.

[hdmjdp]

大佬，你觉得ohem sampler有用吗，能不能跟mosaic一起用？

[zhengye1995]

In my opinion, OHEM generally performs better on datasets with high annotation quality. This dataset contains some annotation noise, which is most likely to have a negative effect.