Closed heiyuxiaokai closed 2 years ago
seem like your misunderstood the random crop. This augmentation replace a random patch of the image with pure noise, rather than replace the whole image with its patch.
seem like your misunderstood the random crop. This augmentation replace a random patch of the image with pure noise, rather than replace the whole image with its patch.
Therefore, the output images of weak strong augmentation own the same object. I understand, and thanks.
Sorry to trouble you again. RandomFlip
in WeakAug
also cause the above situation.
https://github.com/Megvii-BaseDetection/DenseTeacher/blob/ca5c29d04674fdb68c389839a2f8d9ec637ad4c9/coco-p10/augmentations.py#L64
In this code base, student model use StrongAug + Weak Aug while teacher use Weak Aug. Thus the geometric correspondence was retained. see here
In this code base, student model use StrongAug + Weak Aug while teacher use Weak Aug. Thus the geometric correspondence was retained. see here
I see.
Here, teacher model and student model respectively adopt different data augment methods. And the strong one uses
random crop
, which will cause the change in object locations. I'm confused howget_distill_loss
works while dense feature maps are not corresponding. https://github.com/Megvii-BaseDetection/DenseTeacher/blob/ca5c29d04674fdb68c389839a2f8d9ec637ad4c9/coco-p10/runner.py#L242 https://github.com/Megvii-BaseDetection/DenseTeacher/blob/ca5c29d04674fdb68c389839a2f8d9ec637ad4c9/coco-p10/runner.py#L244 https://github.com/Megvii-BaseDetection/DenseTeacher/blob/ca5c29d04674fdb68c389839a2f8d9ec637ad4c9/coco-p10/runner.py#L321