Yet-another-efficientdet-pytorch: data augmentation
I saw that the data augmentation only does left|right flip, I wonder if only one single data augmentation achieves the SoTA or the weights/benchmark metrics reported are achieved by this flip as data augmentation.
Best and thanks for sharing your FAQ, especially the part of your repo vs others.
lolongcovas
All 2 comments
It does. Horizontal flip is the most effective augmentation for a long time. Even the official efficientdet and their paper used hflip and no other augmentation. And lots of classic obj detection/ segmentation networks apply it, like matterport/Mask_RCNN.
Actually I used lots of augmentations before. But when I got desperate, I disabled them. Guess what, the training loss and the validation loss went down really quick. I think augmentations are sometimes and too complex and beyond the model's comprehension.
zylo117
on 9 Apr 2020
thanks for sharing your experience! good to know... :+1:
lolongcovas
on 9 Apr 2020
Related issues
qtw1998
·
3Comments
tetsu-kikuchi
·
3Comments
wangning7149
·
5Comments
mnslarcher
·
5Comments
skylineJLU
·
3Comments
Most helpful comment
It does. Horizontal flip is the most effective augmentation for a long time. Even the official efficientdet and their paper used hflip and no other augmentation. And lots of classic obj detection/ segmentation networks apply it, like matterport/Mask_RCNN.
Actually I used lots of augmentations before. But when I got desperate, I disabled them. Guess what, the training loss and the validation loss went down really quick. I think augmentations are sometimes and too complex and beyond the model's comprehension.